Arcing detection model training method and system based on quantitative perceptual training

Abstract

The invention relates to the technical field of embedded artificial intelligence, and discloses an arc discharge detection model training method and system based on quantitative perception training, and the method comprises the steps: constructing an initial floating point precision neural network model for DC arc discharge fault detection, and defining a plurality of quantization functions for the model to convert floating point parameters; a FakeQuant pseudo quantization strategy is self-defined, a linear calculation layer in an original model is replaced with a quantifiable calculation layer, and a quantization operator is created through the quantifiable calculation layer; performing layer-by-layer differentiation quantitative perception training on the replaced model based on a training data set, and simulating a forward calculation behavior of a low-precision fixed-point type in actual reasoning while keeping floating-point precision back propagation through a quantization operator, so that the weight of the model is adaptive to a quantization error; and converting the trained model parameters into a target low-precision fixed-point type, and generating an end-side deployable arc discharge detection model through model testing and computing power evaluation, thereby realizing high-fidelity conversion from a floating-point model to a low-precision fixed-point model.

Description

Technical Field

[0001] This invention relates to the field of embedded artificial intelligence technology, and in particular to a training method and system for an arc detection model based on quantized perception training. Background Technology

[0002] Deep learning and machine learning training processes primarily involve training a basic model on the server side to obtain a model with satisfactory detection performance, typically using high-precision floating-point numbers as parameters for the arc detection model. In arc detection, the amplitude differences between the normal state and the transient arc state in the current waveform are significant. This leads to substantial differences in the distribution of activation values ​​across different network layers—for example, the input layer needs to process the raw signal with a large dynamic range, while the higher-level feature values ​​are relatively concentrated. Conventional uniform quantization strategies are clearly no longer applicable.

[0003] Meanwhile, the conventional arc fault detection algorithm first selects the model structure, defines the network architecture, trains the model, forces the parameter types to int16 / int8 / int4, and then deploys the model. Directly using floating-point training and then forcibly converting to low precision will cause excessive accuracy loss, especially for sensitive models. Arc fault detection algorithms have high time requirements, needing to ensure high accuracy in judgment at each time step. Secondly, the model structure is small and highly sensitive. Existing technologies directly round the float32 parameters in the model to low precision such as int8 or int16 after training. The model is not adapted to this change, and in arc fault detection, due to the small model size, few parameters, and low redundancy, each parameter has a greater impact on the model output. Small models usually have a simplified structure, and the activation values ​​of model layers are often concentrated. Forcibly using a uniform scaling factor / pruning to compress them will lead to more serious information loss. Due to the training method and the small size of the arc fault detection model, this defect will persist in this field for a long time.

[0004] Solving these problems typically involves adjusting the scaling factor, using int16 instead of the smaller int8, or deploying on a more powerful chip. However, directly deploying it on an edge chip reveals that its runtime is significantly longer than the data arrival time—that is, model inference time is much longer than data acquisition time—thus failing to meet requirements. Detailed analysis shows that the floating-point arithmetic capabilities on edge chips are limited and insufficient to support multiple runs. Edge chips are generally limited by area, power consumption, and cost, and therefore do not have many floating-point units. However, to enhance computing power, they usually provide more int8 or int16 units. Therefore, converting floating-point parameters to fixed-point values, such as int8, is generally considered. While scaling the values ​​according to the model's parameter range allows it to run on an edge chip, the results are unsatisfactory, differing from the offline results during training by about 10%. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the existing technology and provide a training method for arc detection model based on quantization perception training. By defining multi-level quantization functions, injecting pseudo-quantization operators, and performing hierarchical adaptive training, a high-fidelity conversion from floating-point model to low-precision fixed-point model is achieved, thus resolving the contradiction between real-time performance and accuracy in DC arc detection deployed at the end side.

[0006] On the one hand, a training method for an arc detection model based on quantization-aware training is provided, including the following steps: S1: Construct an initial floating-point precision neural network model for DC arcing fault detection, and define multiple quantization functions for the model to transform floating-point parameters; S2: Custom FakeQuant pseudo-quantization strategy, replaces the linear computation layer in the original model with a quantizable computation layer, and creates a quantization operator through the quantizable computation layer; S3: Perform layer-by-layer differential quantization perception training on the replaced model based on the training dataset. By using the quantization operator to maintain the backpropagation of floating-point precision, the forward computation behavior of low-precision fixed-point type in actual inference is simulated, so that the model weights adapt to the quantization error. S4: Convert the trained model parameters into a target low-precision fixed-point type, and generate an end-deployable arc detection model through model testing and computing power evaluation.

[0007] Furthermore, in step S1, defining multiple quantization functions for the model to transform floating-point parameters includes: The quantization function is configurable and is used to dynamically convert the original floating-point parameters into a low-precision integer type while preserving differentiability to support backpropagation. The quantization function's bit width can be dynamically configured to support conversions of different precisions and output quantization.

[0008] Further, in step S2, the creation of the quantization operator includes: The quantizable computation layer integrates quantizable operators, which replace the operators in the original model. During the forward propagation phase, the input data is converted from floating-point to fixed-point and dequantized. During the backpropagation phase, gradient calculations are performed while maintaining floating-point precision.

[0009] Preferably, in step S2, the quantizable computation layer includes a quantized convolutional layer and / or a quantized fully connected layer, and each quantizable computation layer contains independent weight quantization and activation value quantization functions.

[0010] Furthermore, in step S3, the quantization perception training employs a layer-by-layer differentiated quantization strategy, specifically including: S31: Perform quantization sensitivity analysis on different layers of the neural network, and allocate independent quantization bit widths based on the activation value distribution characteristics and computational requirements of each layer; S32: Based on the computational functional requirements obtained from the analysis of each layer, select different scaling factor calculation strategies, including the 3σ principle and quantile truncation; S33: Based on the characteristics of the computing unit of the edge chip, constrain the range of values ​​for the bit width and scaling factor of the quantizable computing layer.

[0011] Preferably, step S3 further includes offline calibration of the trained model, specifically including: S301: Load the model trained by quantization perception, collect current waveform segments from actual application scenarios to generate an unlabeled calibration dataset. The collected current waveform segments need to cover segments in both normal and arcing states. S302: Input the calibration dataset into the trained model, propagate it forward to each quantization layer, and record the dynamic range of the output activation values ​​of each layer. ; S303: Adjust the scaling factor according to the activation value distribution of each layer using the following strategy. : in, Define the target's position width; S304: For asymmetric quantization layers, calculate the zero points. Zero-point offset correction is performed to align the numerical center: in, This is a rounding function; S305: Evaluate the accuracy loss of the optimized injection model on the validation set. If the loss exceeds the preset threshold, iteratively adjust the parameters or switch the quantization layer symmetric / asymmetric strategy.

[0012] Further, in step S4, the model testing specifically includes: Test the baseline arc detection accuracy of the calibrated model on an independent dataset that was not used for training; Add interference, including Gaussian noise and voltage fluctuations, to the input data, and perform robustness tests based on the accuracy degradation rate and fault detection rate. After training and testing, the model input includes critical scenarios such as extremely short electric arcs and weak currents, and boundary condition tests are performed based on the false alarm rate and false negative rate. The consistency test is completed by comparing the output differences of the trained and calibrated model and the initial floating-point precision neural network model on the same input through cosine similarity calculation.

[0013] Preferably, in step S4, the generation of the end-deployable arc detection model further includes: The transformed model is deployed to an embedded edge device, where real-time arc fault detection inference based on low-precision fixed-point arithmetic is performed.

[0014] More preferably, in step S4, the computing power assessment includes: The real-time performance and feasibility of the model inference results on the target edge chip are quantified by using latency testing, multi-task interference testing, and temperature adaptability testing.

[0015] On the other hand, a training system for an arc detection model based on quantization-aware training is provided, including: The model building module is used to build an initial floating-point precision neural network model for DC arcing fault detection, and defines multiple quantization functions for the model to transform floating-point parameters; The quantization operator generation module is used to customize the FakeQuant pseudo-quantization strategy, replace the linear computation layer in the original model with a quantizable computation layer, and create quantization operators through the quantizable computation layer. The hierarchical training module is used to perform layer-by-layer differential quantization-aware training on the replaced model based on the training dataset. The quantization operator simulates the forward computation behavior of low-precision fixed-point type in actual inference while maintaining the backpropagation of floating-point precision, so that the model weights adapt to the quantization error. The model optimization module is used to convert the trained model parameters into a target low-precision fixed-point type, and generate an edge-deployable arc detection model through model testing and computing power evaluation.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention learns about quantization error during training, transforming the model's floating-point weights / activation values ​​into int8 / int4. During backpropagation, the model parameters are automatically adjusted to learn the fluctuations caused by quantization. Ultimately, the model proactively adjusts its weights, making it more robust to operate at low precision and significantly reducing the impact of quantization error on accuracy. This invention finely controls the quantization strategy of each layer during training, allowing different layers to use different quantization strategies, making it more closely match the numerical distribution of each layer and avoiding the overall accuracy being affected by the excessive dynamic range of some layers. This invention simulates the int8 / int4 inference behavior during deployment during training, so its results are consistent with the training effect, accurately predicting the model performance after deployment, resulting in more stable deployment and less debugging. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This invention provides a training method for an arc detection model based on quantized perception training. Figure 2 This is a flowchart illustrating the execution of one embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Quantization-Aware Training (QAT) has the following advantages: It has higher accuracy because the model has been adapted to quantization perturbations during training, so the accuracy after deployment is close to that of float32.

[0020] The inference consistency is good because the same parameters are used for training and inference without any cropping or scaling, so the results are consistent.

[0021] It supports flexible strategies, such as channel-by-channel quantization and asymmetric quantization, to better integrate the model with the performance of the edge chip.

[0022] The specific embodiments of the present invention will be described below with reference to the accompanying drawings and examples.

[0023] Example 1 Please see Figure 1 The technical solution for training an arc detection model based on quantization-perception training provided in this embodiment includes the following steps: S1: Construct an initial floating-point precision neural network model for DC arcing fault detection, and define multiple quantization functions for the model to transform floating-point parameters; S2: Custom FakeQuant pseudo-quantization strategy, replaces the linear computation layer in the original model with a quantizable computation layer, and creates a quantization operator through the quantizable computation layer; S3: Perform layer-by-layer differential quantization-aware training (QAT) on the replaced model based on the training dataset. By using the quantization operator to maintain the backpropagation of floating-point precision, the forward computation behavior of low-precision fixed-point type in actual inference is simulated, so that the model weights adapt to the quantization error. S4: Convert the trained model parameters into a target low-precision fixed-point type, and generate an end-deployable arc detection model through model testing and computing power evaluation.

[0024] This technical solution converts floating-point weights / activation values ​​to int8 / int4. During backpropagation, the model parameters are automatically adjusted to learn the fluctuations caused by quantization. Ultimately, the model proactively adjusts the weights, making it more robust to operate at low precision and significantly reducing the impact of quantization errors on accuracy.

[0025] Specifically, in step S1, an extensible family of quantization functions is designed to achieve lossless conversion of floating-point parameters to fixed-point types such as int16 / int8 / int4, resolving the conflicting requirements of a wide dynamic range of current (±100A) and high fault pulse resolution (<1A) in arc detection. Further defining multiple quantization functions for the model to convert floating-point parameters includes: The quantization function is configurable and is used to dynamically convert the original floating-point parameters into a low-precision integer type while preserving differentiability to support backpropagation. The quantization function's bit width can be dynamically configured to support conversions of different precisions and output quantization.

[0026] In this embodiment, for fault pulses, we first separate the positive and negative pulse parameters and process them independently to avoid the normal current overwhelming the fault pulse. The quantization function needs to include two key parameters: a scaling factor and a zero point, expressed as follows: in, For the input floating-point value, This represents the quantized integer output. This represents the result of aligning floating-point zeros to integer zeros. In this embodiment, the integer constraints are as follows: int8: , ; int4: , ; Int16: , .

[0027] Furthermore, the formula used to simulate inverse quantization error is as follows: in, This indicates the mapping ratio from floating-point values ​​to integers.

[0028] Next, the quantization operator replacement in step S2 is performed, which further includes: The quantizable computation layer integrates quantizable operators, which replace the operators in the original model. During the forward propagation phase, the input data is converted from floating-point to fixed-point and dequantized. During the backpropagation phase, gradient calculations are performed while maintaining floating-point precision.

[0029] More preferably, the quantizable computation layer includes quantized convolutional layers and / or quantized fully connected layers, and each quantizable computation layer contains independent weight quantization and activation value quantization functions.

[0030] After completing the above preparations, the model undergoes quantization-aware training, which employs a layer-by-layer differentiated quantization strategy, specifically including: S31: Perform quantization sensitivity analysis on different layers of the neural network, and allocate independent quantization bit widths based on the activation value distribution characteristics and computational requirements of each layer; S32: Based on the computational functional requirements obtained from the analysis of each layer, select different scaling factor calculation strategies, including the 3σ principle and quantile truncation; S33: Based on the characteristics of the computing unit of the edge chip, constrain the range of values ​​for the bit width and scaling factor of the quantizable computing layer.

[0031] This embodiment allows for fine-grained control of the quantization strategy for each layer during training. Different layers can use different quantization strategies to better match the numerical distribution of each layer, thus avoiding the impact of excessive dynamic range on the overall accuracy of certain layers.

[0032] Generally, the amplitude of the input convolutional layer current waveform changes drastically, so the configuration principle is high bit width (16 bits) plus dynamic scaling factor; the arc features of the intermediate feature layer are abstracted and concentrated, so the configuration principle is medium bit width (8 bits) plus symmetric quantization; the output classification layer is used for probability output and requires high resolution, so the configuration principle is customized bit width (10-12 bits) plus asymmetric quantization.

[0033] Specifically, in this embodiment, step S32, selecting different scaling factor calculation strategies, includes: For the Gaussian distribution layer, a scaling factor calculation strategy based on standard deviation is adopted; for the long-tailed distribution layer, a scaling factor calculation strategy based on quantile truncation is adopted.

[0034] In summary, during training in step S3, the behavior of low-precision quantization such as int8 during actual inference is simulated, but the backpropagation capability of float32 is still maintained. That is, the int8 / int4 inference behavior during deployment is simulated. Therefore, the results are consistent with the training effect, and the model performance after deployment can be accurately predicted, resulting in more stable deployment and less debugging.

[0035] After training is complete, the trained model is calibrated offline, specifically including: S301: Load the model trained by quantization perception, collect current waveform segments from actual application scenarios to generate an unlabeled calibration dataset. The collected current waveform segments need to cover segments in both normal and arcing states. S302: Input the calibration dataset into the trained model, propagate it forward to each quantization layer, and record the dynamic range of the output activation values ​​of each layer. ; S303: Adjust the scaling factor according to the activation value distribution of each layer using the following strategy. : In this embodiment, a truncation threshold is used to limit the dynamic range of the long-tailed distribution layer, as shown below: in, Define the target location width. The cutoff coefficient is adjustable, preserving the core distribution range; S304: For asymmetric quantization layers, calculate the zero points. Zero-point offset correction is performed to align the numerical center: in, This is a rounding function; S305: Evaluate the accuracy loss of the optimized injection model on the validation set. If the loss exceeds the preset threshold, iteratively adjust the parameters or switch the quantization layer symmetric / asymmetric strategy.

[0036] Finally, step S4 is performed, including testing, computing power evaluation, and model deployment. Specifically, the model testing includes: Test the baseline arc detection accuracy of the calibrated model on an independent dataset that was not used for training; Add interference, including Gaussian noise and voltage fluctuations, to the input data, and perform robustness tests based on the accuracy degradation rate and fault detection rate. After training and testing, the model input includes critical scenarios such as extremely short electric arcs and weak currents, and boundary condition tests are performed based on the false alarm rate and false negative rate. The consistency test is completed by comparing the output differences of the trained and calibrated model and the initial floating-point precision neural network model on the same input through cosine similarity calculation.

[0037] In this embodiment, we use dynamic threshold calibration technology to automatically adjust the arc judgment threshold according to the real-time noise level; we construct a type-specific confusion matrix by testing different arc types separately to perform fault mode separation testing; and we verify the timing continuity by detecting the output stability of 10 consecutive sampling points.

[0038] The arc detection model that can be deployed on the generating end side further includes: The transformed model is deployed to an embedded edge device, where real-time arc fault detection inference based on low-precision fixed-point arithmetic is performed.

[0039] The computing power assessment includes: The real-time performance and feasibility of the model inference results on the target edge chip are quantified by using latency testing, multi-task interference testing, and temperature adaptability testing.

[0040] In summary, the specific implementation process of this technical solution is as follows: Figure 2 As shown, it includes: Step 1: Define multiple quantization functions, which can convert floating-point parameters into types such as int16, int8, and int4; Step 2: Create a custom FakeQuant module to simulate the behavior of low-precision quantization such as int8 during actual inference, but still maintain the backpropagation capability of float32; Step 3: Create quantifiable operators, such as linear and CNN2D, to replace the operators in the original model; Step 4: Quantization-aware training. Based on the quantization operator, complete the training using the normal PyTorch training process. Step 5: Testing – Test the unknown data to evaluate the accuracy. Step Six: Computing power assessment, assessing the feasibility of model inference results and real-time performance on the target chip.

[0041] On the other hand, this embodiment also provides a training system for an arc detection model based on quantization-aware training, including: The model building module is used to build an initial floating-point precision neural network model for DC arcing fault detection, and defines multiple quantization functions for the model to transform floating-point parameters; The quantization operator generation module is used to customize the FakeQuant pseudo-quantization strategy, replace the linear computation layer in the original model with a quantizable computation layer, and create quantization operators through the quantizable computation layer. The hierarchical training module is used to perform layer-by-layer differential quantization-aware training on the replaced model based on the training dataset. The quantization operator simulates the forward computation behavior of low-precision fixed-point type in actual inference while maintaining the backpropagation of floating-point precision, so that the model weights adapt to the quantization error. The model optimization module is used to convert the trained model parameters into a target low-precision fixed-point type, and generate an edge-deployable arc detection model through model testing and computing power evaluation.

[0042] It should be noted that the steps in the arc detection model training method based on quantization perception training provided in this embodiment can be implemented based on the corresponding modules in the arc detection model training system based on quantization perception training. Those skilled in the art can refer to the technical solution of the system to implement the steps of the method. That is, the embodiments in the system can be understood as preferred examples of implementing the method, and will not be elaborated here.

[0043] Besides implementing the system and its various devices provided by this invention in purely computer-readable program code, the same functions can be achieved by logically programming the method steps, making the system and its various devices of this invention appear as logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices provided by this invention can be considered as a hardware component, and the devices included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0044] Finally, it should be noted that the above description is only a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be pointed out that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

[0045] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A training method for an arc detection model based on quantization-perception training, characterized in that, Includes the following steps: S1: Construct an initial floating-point precision neural network model for DC arcing fault detection, and define multiple quantization functions for the model to transform floating-point parameters; S2: Custom FakeQuant pseudo-quantization strategy, replaces the linear computation layer in the original model with a quantizable computation layer, and creates a quantization operator through the quantizable computation layer; S3: Perform layer-by-layer differential quantization perception training on the replaced model based on the training dataset. By using the quantization operator to maintain the backpropagation of floating-point precision, the forward computation behavior of low-precision fixed-point type in actual inference is simulated, so that the model weights adapt to the quantization error. S4: Convert the trained model parameters into a target low-precision fixed-point type, and generate an end-deployable arc detection model through model testing and computing power evaluation.

2. The arc detection model training method based on quantization-perception training according to claim 1, characterized in that, In step S1, defining multiple quantization functions for the model to transform floating-point parameters further includes: The quantization function is configurable and is used to dynamically convert the original floating-point parameters into a low-precision integer type while preserving differentiability to support backpropagation. The quantization function's bit width can be dynamically configured to support conversions of different precisions and output quantization.

3. The arc detection model training method based on quantization-perception training according to claim 1, characterized in that, In step S2, the creation of the quantization operator further includes: The quantizable computation layer integrates quantizable operators, which replace the operators in the original model. During the forward propagation phase, the input data is converted from floating-point to fixed-point and dequantized. During the backpropagation phase, gradient calculations are performed while maintaining floating-point precision.

4. The arc detection model training method based on quantization-perception training according to claim 1, characterized in that, In step S2, the quantizable computation layer includes a quantized convolutional layer and / or a quantized fully connected layer, and each quantizable computation layer contains independent weight quantization and activation value quantization functions.

5. The arc detection model training method based on quantization-perception training according to claim 1, characterized in that, In step S3, the quantization perception training employs a layer-by-layer differential quantization strategy, specifically including: S31: Perform quantization sensitivity analysis on different layers of the neural network, and allocate independent quantization bit widths based on the activation value distribution characteristics and computational requirements of each layer; S32: Based on the computational functional requirements obtained from the analysis of each layer, select different scaling factor calculation strategies, including the 3σ principle and quantile truncation; S33: Based on the characteristics of the computing unit of the edge chip, constrain the range of values ​​for the bit width and scaling factor of the quantizable computing layer.

6. The arc detection model training method based on quantization-perception training according to claim 1, characterized in that, Step S3 also includes offline calibration of the trained model, specifically including: S301: Load the model trained by quantization perception, collect current waveform segments from actual application scenarios to generate an unlabeled calibration dataset. The collected current waveform segments need to cover segments in both normal and arcing states. S302: Input the calibration dataset into the trained model, propagate it forward to each quantization layer, and record the dynamic range of the output activation values ​​of each layer. ; S303: Adjust the scaling factor according to the activation value distribution of each layer using the following strategy. : in, Define the target's position width; S304: For asymmetric quantization layers, calculate the zero points. Zero-point offset correction is performed to align the numerical center: in, This is a rounding function; S305: Evaluate the accuracy loss of the optimized injection model on the validation set. If the loss exceeds the preset threshold, iteratively adjust the parameters or switch the quantization layer symmetric / asymmetric strategy.

7. The arc detection model training method based on quantization-perception training according to claim 6, characterized in that, In step S4, the model testing specifically includes: Test the baseline arc detection accuracy of the calibrated model on an independent dataset that was not used for training; Add interference, including Gaussian noise and voltage fluctuations, to the input data, and perform robustness tests based on the accuracy degradation rate and fault detection rate. After training and testing, the model input includes critical scenarios such as extremely short electric arcs and weak currents, and boundary condition tests are performed based on the false alarm rate and false negative rate. The consistency test is completed by comparing the output differences of the trained and calibrated model and the initial floating-point precision neural network model on the same input through cosine similarity calculation.

8. The arc detection model training method based on quantization-perception training according to claim 1, characterized in that, In step S4, the generation of the end-deployable arc detection model further includes: The transformed model is deployed to an embedded edge device, where real-time arc fault detection inference based on low-precision fixed-point arithmetic is performed.

9. The arc detection model training method based on quantization-perception training according to claim 8, characterized in that, In step S4, the computing power assessment includes: The real-time performance and feasibility of the model inference results on the target edge chip are quantified by using latency testing, multi-task interference testing, and temperature adaptability testing.

10. A training system for an arc detection model based on quantization-perception training, characterized in that, include: The model building module is used to build an initial floating-point precision neural network model for DC arcing fault detection, and defines multiple quantization functions for the model to transform floating-point parameters; The quantization operator generation module is used to customize the FakeQuant pseudo-quantization strategy, replace the linear computation layer in the original model with a quantizable computation layer, and create quantization operators through the quantizable computation layer. The hierarchical training module is used to perform layer-by-layer differential quantization-aware training on the replaced model based on the training dataset. The quantization operator simulates the forward computation behavior of low-precision fixed-point type in actual inference while maintaining the backpropagation of floating-point precision, so that the model weights adapt to the quantization error. The model optimization module is used to convert the trained model parameters into a target low-precision fixed-point type, and generate an edge-deployable arc detection model through model testing and computing power evaluation.

Images