Risk perception neural network quantification method and device for aero-engine health management

Abstract

The embodiment of the invention provides a risk awareness neural network quantification method and device for aero-engine health management, and relates to the technical field of neural networks, and the method comprises the following steps: obtaining a floating point neural network model and a representative data set, aiming at each expert unit, based on the representative data set, obtaining a floating point neural network model; calculating and fusing the sensitivity measurement and the dynamic range measurement of the expert unit, and generating an importance score of the expert unit; on the basis of a preset resource constraint condition and a candidate bit width set, according to the importance score of the expert unit, distributing a discrete target quantization bit width for the expert unit; and inserting a false quantization operator corresponding to the target quantization bit width distributed by the expert unit, generating an initial mixed bit width model, performing quantization perception adjustment on the initial mixed bit width model based on the task data set, and generating a risk perception neural network. According to the scheme, efficient and reliable deployment of the neural network model is realized.

Description

Technical Field

[0001] This invention relates to the field of neural network technology, and in particular to a method and apparatus for quantifying risk perception neural networks for aircraft engine health management. Background Technology

[0002] In resource-constrained scenarios such as aero-engine health management, edge intelligence, and real-time control, the deployment of deep learning models faces significant challenges in terms of computing power, power consumption, and storage. Model quantization technology, by compressing parameters and activation values ​​represented as floating-point numbers in the network to low-bit integers (such as INT8, INT4, etc.), can significantly reduce the model's memory footprint and inference latency, and is a key technology for achieving efficient deployment of models on resource-constrained devices.

[0003] However, the quantization process introduces computational errors, potentially altering the network's numerical stability and output distribution. This is especially true in risk-sensitive tasks, such as predicting the remaining service life of aero-engines, which require not only accurate predictions but also precise characterization of risk structures like tail risks and decision biases. In such cases, even minor quantization perturbations can be amplified, leading to misjudgments of critical risk events and posing serious safety hazards.

[0004] Existing quantization methods mainly fall into two categories. The first is uniform bit-width quantization, which applies the same quantization bit width to all layers or modules of the network. This "one-size-fits-all" approach ignores the varying sensitivities of different units in the network to quantization errors, potentially leading to over-quantization of critical units and a significant degrade in model performance. Simultaneously, it fails to adequately quantize redundant units, wasting valuable computational resources. The second category is hybrid bit-width quantization based on experience or fixed rules, where a higher bit width is manually assigned to a few critical layers. This type of method heavily relies on expert experience, lacks data-driven automation and systematic measurement, is difficult to generalize to different network structures and tasks, and cannot guarantee an optimal performance trade-off within a given resource budget. Therefore, existing technologies generally suffer from the inability to adaptively and finely allocate bit widths based on the importance of different network components, making it difficult to achieve an effective balance between model compression efficiency, prediction accuracy, and risk structure preservation. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a risk perception neural network quantization method for aero-engine health management, to solve the technical problem that existing neural network quantization methods are unable to adaptively allocate bit width according to the importance of each part of the network, thus failing to achieve an optimal balance between model compression efficiency, prediction accuracy, and risk structure preservation. The method includes:

[0006] Obtain a pre-trained floating-point neural network model and a representative dataset for calibration. For each expert unit in the floating-point neural network, calculate the sensitivity metric and dynamic range metric of the expert unit based on the representative dataset. Fuse the sensitivity metric and dynamic range metric corresponding to each expert unit to generate the importance score of the expert unit. Based on preset resource constraints and a candidate bit width set, a discrete target quantization bit width is allocated to the expert unit according to the importance score of the expert unit, wherein the level of the importance score is proportional to the level of the discrete target quantization bit width. In each expert unit of the floating-point neural network model, a pseudo-quantization operator corresponding to the target quantization bit width assigned to the expert unit is inserted to generate an initial mixed bit width model. The initial mixed bit width model is then subjected to quantization-aware adjustment based on the task dataset. Furthermore, weights related to the importance score of the expert unit are introduced into the loss function of the initial mixed bit width model to generate the final mixed bit width quantization model, which serves as the risk-aware neural network.

[0007] This invention also provides a risk-aware neural network quantization device for aero-engine health management, addressing the technical problem in existing network quantization methods that struggle to adaptively allocate bit widths based on the importance of different network components, thus failing to achieve an optimal balance between model compression efficiency, prediction accuracy, and risk structure preservation. The device includes: The importance calculation module is used to obtain a pre-trained floating-point neural network model and a representative dataset for calibration. For each expert unit in the floating-point neural network, based on the representative dataset, the module calculates the sensitivity metric and dynamic range metric of the expert unit. The module then fuses the sensitivity metric and dynamic range metric corresponding to each expert unit to generate the importance score of the expert unit. The discrete target quantization bit width allocation module is used to allocate discrete target quantization bit width to the expert unit based on preset resource constraints and a candidate bit width set, according to the importance score of the expert unit, wherein the level of the allocated importance score is proportional to the level of the discrete target quantization bit width. A risk-aware neural network construction module is used to insert a pseudo-quantization operator corresponding to the target quantization bit width assigned to the expert unit in each expert unit of the floating-point neural network model to generate an initial mixed bit width model, perform quantization-aware adjustment on the initial mixed bit width model based on the task dataset, and introduce weights related to the importance score of the expert unit into the loss function of the initial mixed bit width model to generate a final mixed bit width quantization model, which serves as the risk-aware neural network.

[0008] Compared with the prior art, the beneficial effects that at least one technical solution adopted in the embodiments of this specification can achieve include at least: The embodiments of the present invention enable the efficient and reliable deployment of neural network models while meeting stringent aerospace application requirements. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of a risk perception neural network quantification method for aero-engine health management provided by an embodiment of the present invention; Figure 2 This is a structural block diagram of a risk perception neural network quantification device for aircraft engine health management provided in an embodiment of the present invention. Detailed Implementation

[0011] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0012] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] In this embodiment of the invention, a risk perception neural network quantification method for aero-engine health management is provided, such as... Figure 1 As shown, the method includes: Step S101: Obtain a pre-trained floating-point neural network model and a representative dataset for calibration. For each expert unit in the floating-point neural network, calculate the sensitivity metric and dynamic range metric of the expert unit based on the representative dataset. Fuse the sensitivity metric and dynamic range metric corresponding to each expert unit to generate the importance score of the expert unit. Step S102: Based on preset resource constraints and candidate bit width set, allocate discrete target quantization bit width to the expert unit according to the importance score of the expert unit, wherein the level of the importance score allocation is proportional to the level of the discrete target quantization bit width; Step S103: In each expert unit of the floating-point neural network model, insert a pseudo-quantization operator corresponding to the target quantization bit width assigned to the expert unit to generate an initial mixed bit width model. Perform quantization-aware adjustment on the initial mixed bit width model based on the task dataset, and introduce weights related to the importance score of the expert unit into the loss function of the initial mixed bit width model to generate the final mixed bit width quantization model, which serves as the risk-aware neural network.

[0014] In specific implementation, the following steps are used to calculate the sensitivity metric and dynamic range metric of each expert unit in the floating-point neural network based on the representative dataset: Using the representative dataset, a forward propagation process is performed on the floating-point neural network model; for each expert unit, the variance of the output activation values ​​on the representative dataset is calculated and used as a measure of the dynamic range of the expert unit. ,in, The output activation value of the expert unit; using the representative dataset, perform a backpropagation process on the floating-point neural network model; for each expert unit, collect the k-th... Trainable parameters of each expert unit The gradient value is calculated, and the expected value of the square of the gradient value on the representative dataset is used as a sensitivity measure of the expert unit. ,in, As expected, For partial differential equations, is the loss function of the floating-point neural network model.

[0015] In specific implementation, the following steps are used to fuse the sensitivity metric and the dynamic range metric corresponding to each expert unit to generate the importance score of the expert unit: The sensitivity measure for each expert unit and the dynamic range metric Scale standardization is performed separately, and the standardized sensitivity metric and the standardized dynamic range metric are fused together to obtain the original importance score of the expert unit. The original importance scores of all expert units are normalized to generate the final importance score of each expert unit.

[0016] In specific implementation, the sensitivity measurement for each expert unit is achieved through the following steps. and the dynamic range metric Scale standardization is performed separately, and the standardized sensitivity metric and the standardized dynamic range metric are then fused to obtain the original importance score of the expert unit: For the sensitivity measure and the dynamic range metric Perform logarithmic transformations on each sample to generate a standardized sensitivity metric. and the dynamic range metric after standardization ,in, , ε is a preset minimum normal number; weighted linear fusion is used to calculate the original importance score of the expert unit. , where α and β are preset fusion weight coefficients.

[0017] In specific implementation, the following steps are used to allocate discrete target quantization bit widths to the expert units based on preset resource constraints and candidate bit width sets, according to the importance scores of the expert units: Based on the importance score of each expert unit The expert unit Mapped to continuous bit width requirement values ,in, This is the upper bound of the preset bit width requirement. Set a lower bound for the preset bit width requirement; calculate the total resource consumption under the current bit width allocation scheme. When the resource constraint is a storage size constraint, the total resource consumption under the current bit width allocation scheme is... , For the first The number of parameters to be quantized in each expert unit, and the total resource consumption under the current bit-width allocation scheme when the resource constraint is an inference delay constraint. , This is a lookup table function based on the pre-calibrated bit width to inference delay of the target hardware platform; under constraints... Next, an optimization objective function is constructed, which is to maximize the total weighted importance score. ,in, The preset resource constraint threshold, This is a resource consumption function. The objective function is a monotonically non-decreasing function with respect to bit width; the bit width requirement is adjusted through the optimization objective function. To satisfy Furthermore, the bit width is optimized until a bit width requirement value that satisfies the aforementioned resource constraints is obtained. The bit width requirement that satisfies the aforementioned resource constraints. Projected onto the candidate bit-width set In this process, the discrete target quantization bit width of each expert unit is obtained. .

[0018] In specific implementation, the following steps are used to insert a pseudo-quantization operator corresponding to the target quantization bit width allocated to the expert unit in each expert unit of the floating-point neural network model, thereby generating an initial mixed bit width model: For each expert unit Quantization bit width based on the allocated discrete target Determine the expert unit The range of integer values Among them, the minimum value maximum value ; refers to the expert unit activation value Constructing a pseudo-quantization operator ,in, For learnable quantization step size parameters, To quantize the zero-point parameters, This is a truncation function. The dummy quantization operator is inserted before the weights and / or activation paths of each expert unit in the floating-point neural network model to replace the original floating-point operations, thus obtaining the initial mixed bit-width model.

[0019] In specific implementation, the initial mixed bit-width model is quantized and adjusted based on the task dataset through the following steps, and weights related to the importance scores of the expert units are introduced into the loss function of the initial mixed bit-width model to generate the final mixed bit-width quantization model: Construct a weighted loss function for quantizing perception adjustment. ,in, The total number of expert units. For calculation based on the task dataset and related to the first Each expert unit outputs the relevant original task loss components. In order to be with the first The final importance score of each expert unit Positively correlated expert perception weights The regularization coefficient is . To quantify the regularization term; based on the final importance score of each expert unit. Set the expert perception weight ,in, For temperature hyperparameters; using the task dataset, based on the expert-perceived weights. The constructed weighted loss function To achieve this, the initial mixed bit-width model is trained iteratively for multiple rounds until a preset stopping condition is met, thereby obtaining the final mixed bit-width quantization model.

[0020] In practice, the quantization regularization term is constructed through the following steps: The quantization regularization term For weight quantization error regularization term Activate the distribution alignment regularization term Gradient Preservation Regularization Term One or more combinations thereof; weight quantization error regularization term ,in, For the first The original floating-point weights of each expert unit, For pseudo-quantization operators, For learnable quantization step size parameters, To quantize the zero-point parameters; activate the distribution alignment regularization term. ,in, For the first The set of activation values ​​for each expert unit on the task dataset. For probability distribution, Kullback-Leibler divergence; gradient-preserving regularization term ,in, For the first The gradient of the task loss for each expert unit in the floating-point model. This represents the gradient of the task loss in the quantization model.

[0021] In practice, the floating-point neural network model is constructed through the following steps: The floating-point neural network model is used as a risk perception network for predicting remaining useful life. The floating-point neural network model includes a shared feature encoder and at least one risk quantization output head. The risk quantization output head is an independent expert unit and is configured to output a conditional quantization estimate corresponding to a preset risk level. When allocating discrete target quantization bit width to the expert unit, the target quantization bit width allocated to the risk quantization output head is set to a value not lower than a preset bit width protection threshold.

[0022] The nouns and terms used in the embodiments of this application are explained, and the nouns and terms used in the embodiments of this application are subject to the following interpretations.

[0023] (1) Quantization: This refers to a model compression technique. Its core idea is to map the weight parameters or activation values ​​in the neural network model, which are represented by high-precision floating-point numbers (such as 32-bit or 16-bit floating-point numbers), to low-bit-width integers (such as 8-bit, 6-bit, or 4-bit integers). This process aims to significantly reduce the model's storage space occupation, memory bandwidth requirements, and arithmetic complexity in the computation process, thereby accelerating the inference speed of the model on resource-constrained hardware platforms.

[0024] (2) Mixed-Bitwidth: This refers to a non-uniform quantization strategy. Unlike schemes that use a uniform bitwidth for all layers or modules of the entire network, mixed-bitwidth allows different parts of the network (e.g., different layers, channels, or specific tensors) to use different quantization bitwidths. The basic idea is to allocate higher bitwidths to parts of the model that are more sensitive to accuracy, and to use lower bitwidths for parts with higher redundancy, in order to achieve a better balance between model performance and compression ratio.

[0025] (3) Expert Unit: In the context of this application, an expert unit is a functional division of the internal structure of a neural network model. It can refer to a basic computational unit in the network, such as a single convolutional layer or a fully connected layer; or it can be a more complex structure, such as a residual block, an attention sub-layer, or an output head module specifically designed for a particular sub-task. The purpose of dividing the network into expert units is to enable the independent evaluation and processing of the contribution and sensitivity of each unit to the overall task.

[0026] (4) Fisher Information (F): In information geometry and statistics, Fisher information is a measure of the information contained in a random variable about its model parameters. In the practice of neural networks, its approximate form, namely the second moment of the gradient of the loss function with respect to the model parameters (i.e., the expectation of the square of the gradient value), is usually used to quantify the sensitivity of a parameter or a set of parameters (such as the parameters of an expert unit) to the final output of the model (measured by the loss function). The larger the Fisher information value, the greater the impact of a small change in the parameter on the model loss, that is, the more important the parameter.

[0027] (5) Activation Variance (V): This refers to the degree of variation or dispersion of the output activation value of an expert unit (such as a network layer) when it passes through a representative dataset. A larger activation variance usually means that the output of the unit can encode richer and more diverse information, has a wider dynamic range, and responds more significantly to changes in the input. Therefore, activation variance can be used as an indicator to measure the amount of representational information carried by the expert unit.

[0028] (6) Importance Score (I): This refers to a comprehensive metric proposed in this application, used to evaluate the relative importance of each expert unit in a neural network. The score is generated by fusing the sensitivity metric (e.g., Fisher information) and dynamic range metric (e.g., activation variance) of the expert unit. Expert units with high importance scores are considered key components of the model and should be handled more carefully during quantization, for example, by assigning them a higher bit width.

[0029] This invention provides a risk perception neural network quantization method for aero-engine health management. Through an expert-perceived hybrid bit-width quantization strategy, it achieves effective model compression while ensuring key model performance.

[0030] The core process of this method comprises three main stages. First, a pre-trained floating-point neural network model is "portrayed" to determine the importance of its different "expert units." Second, based on this importance portrait and given hardware resource constraints, an optimal quantization bit width is dynamically allocated to each expert unit. Finally, based on the allocated hybrid bit width scheme, the model is fine-tuned with quantization awareness to recover the accuracy loss introduced by quantization, ultimately generating a deployable and efficient hybrid bit width quantization model.

[0031] Specifically, the first step, S101, involves acquiring a pre-trained floating-point neural network model for aero-engine health management (such as remaining service life prediction) and a small, representative dataset for calibration. This dataset should reflect typical data distributions that the model might encounter in real-world applications. For each defined expert unit in the floating-point neural network model, the method calculates two key metrics based on this representative dataset: a sensitivity metric and a dynamic range metric.

[0032] Sensitivity metrics measure the impact of parameter changes on the overall model performance. An expert unit is considered highly sensitive if its parameters significantly affect the final loss function. Dynamic range metrics measure the range of activation values ​​and information richness of the expert unit's output. A wide dynamic range indicates that an expert unit can express richer feature information. These two metrics reflect the importance of the expert unit from different perspectives.

[0033] After calculating the sensitivity metric and dynamic range metric for each expert unit, step S101 further includes fusing these two metrics to generate a single, comprehensive importance score for each expert unit. This score quantitatively represents the importance of the expert unit within the entire network. The fusion process aims to combine parameter sensitivity and information representation capabilities to more comprehensively assess the potential impact of quantification on the unit.

[0034] Next, the method proceeds to step S102, the dynamic bit-width allocation stage. The core task of this stage is to allocate an appropriate quantization bit width to each expert unit based on the expert unit importance score generated in the previous step. This allocation process is guided by two main factors: first, pre-defined hardware resource constraints, such as the model's final total storage size not exceeding a threshold, or the model's inference latency on the target hardware must be below a certain time limit; and second, a set of candidate bit widths to choose from, such as hardware support for 4-bit, 6-bit, and 8-bit integer operations.

[0035] The basic principle of allocation is that the importance score of an expert unit is proportional to the target quantization bit width it is assigned. That is, expert units with higher importance scores will be assigned a higher bit width (e.g., 8 bits) to more accurately retain their computational results and reduce quantization errors. Conversely, expert units with lower importance scores can be assigned a lower bit width (e.g., 4 bits) to maximize the conservation of storage and computational resources. This differentiated allocation strategy allows for the maximum retention of information crucial to model performance while meeting overall resource constraints.

[0036] After determining the target quantization bit width for each expert unit, the method proceeds to step S103, the model construction and fine-tuning stage. First, in the original floating-point neural network model structure, a "dummy quantization operator" is inserted for each expert unit. This operator simulates the error caused by real quantization operations (i.e., the process from floating-point to low-bit integer and back to floating-point) during training. Since each expert unit is assigned a different bit width, the inserted dummy quantization operator corresponds to the target quantization bit width of that unit, thus constructing an initial mixed bit width model.

[0037] Subsequently, this initial mixed bit-width model needs to undergo quantization-aware tuning (also known as quantization-aware fine-tuning). This process involves training the model for several additional rounds using a task dataset (which can be the original training dataset or a subset thereof). The goal is to adapt the model's parameters (weights) to the noise and errors introduced by the quantization operation, thereby restoring performance. A key improvement lies in the specially designed loss function at this stage: when calculating the total loss, weights related to the importance scores of expert units are introduced. Specifically, expert units with higher importance scores are given greater weight for their contribution to the total loss. This means that during training, the optimization algorithm will focus more on reducing the errors of these key expert units.

[0038] After multiple rounds of quantization-aware adjustments, the training process ends when preset stopping conditions are met (e.g., reaching a certain number of training epochs or model performance convergence). The resulting model is the final hybrid bit-width quantization model. This model is not only smaller and faster inference, but also maintains its performance effectively in critical tasks such as aero-engine risk perception due to the adoption of expert-aware bit-width allocation and weighted fine-tuning strategies. Ultimately, this model can be deployed as a risk perception neural network for use in practical aero-engine health management systems.

[0039] The overall logic of this method embodies a resource optimization concept. Instead of treating all network parts equally, it identifies key components through precise "profile" analysis and prioritizes allocating the limited bit budget to them. This strategy makes the model compression process more intelligent and efficient.

[0040] In terms of interaction, the "importance score" generated in step S101 is the core of the entire method. On the one hand, it directly guides the bit width allocation decision in step S102, determining the "bit budget" of each expert unit; on the other hand, it is used to construct the weighted loss function in step S103, guiding the direction and focus of model fine-tuning. The discrete bit width of each expert unit output in step S102 defines the specific structure of the initial mixed bit width model in step S103.

[0041] In principle, this method is based on a fundamental assumption: the contributions of different parts of a neural network to the final performance are uneven. By combining the sensitivity of parameters to loss (Fisher information) and the information content of layer outputs (activation variance), it is possible to effectively identify those "expert units" that are crucial to model performance and numerical stability. Maintaining high numerical accuracy for these units while applying greater compression to other redundant or insensitive units is key to achieving efficient compression without excessively sacrificing performance.

[0042] Furthermore, the expert-perceived weighting mechanism in the quantization-aware fine-tuning process is equivalent to providing a "priority protection" instruction for the model's learning process. The optimizer is guided to prioritize correcting or compensating for errors caused by quantization in important units, which further strengthens the stability of key functions, especially in risky quantization tasks that require precise numerical output.

[0043] In summary, the basic implementation method achieves automated and differentiated quantization of neural networks through a data-driven approach. Its technical advantage lies in the fact that, compared to traditional uniform bit-width quantization methods, this method can better maintain the prediction accuracy and risk structure integrity of the aero-engine risk perception model while meeting strict hardware resource constraints (such as storage and latency), thus improving the reliability and deployment feasibility of the quantization model.

[0044] In a preferred embodiment, the process of calculating the sensitivity and dynamic range metrics of the expert units is specified. Based on the above embodiments, to obtain reliable metrics, it is first necessary to perform a complete forward propagation process on the pre-trained floating-point neural network model using the representative dataset. During this process, the output activation values ​​generated by each expert unit when processing data can be collected.

[0045] For each expert unit Its dynamic range measurement Defined as its output activation value The variance over the entire representative dataset. The specific formula for calculation is:

[0046] in, Representing the The set of output activation values ​​of each expert unit This indicates the operation of calculating variance. This metric reflects the signal strength and range of variation of the unit's output. The larger the variance, the wider the dynamic range of the unit's output, and the more information it may carry.

[0047] Next, to calculate the sensitivity metric, a backpropagation process needs to be performed on the floating-point neural network model using the same representative dataset. During backpropagation, the model loss function can be calculated. Gradients for all trainable parameters in the network. Gradients for each expert unit. Collect all trainable parameters within it. gradient value Sensitivity measurement ,in, This represents the operation of calculating the expected value. The symbol is for partial differentials. The loss function for the floating-point neural network model. Let be the trainable parameters of the i-th expert unit. This metric is a valid approximation of the diagonal elements of the Fisher information matrix in practice, and it measures how sensitive the expert unit's parameters are to the model's final loss. The larger the value, the more critical the impact of that unit on the model performance.

[0048] The above scheme provides a concrete, operable, and theoretically reasonable calculation method for the abstract "sensitivity metric" and "dynamic range metric." Through one forward propagation and one backward propagation, these two core metrics can be efficiently calculated for all expert units in the network, providing a solid data foundation for subsequent importance assessment and bit width allocation. This gradient and activation statistics-based method is more objective and reliable than methods based on experience or heuristic rules.

[0049] Furthermore, based on the above embodiments, the specific process of fusing the sensitivity metric and dynamic range metric of each expert unit to generate an importance score may include the following steps. First, due to the sensitivity metric... and dynamic range measurement The numerical scales of the two metrics may differ significantly, and direct fusion could lead to one metric dominating the computation. Therefore, it is necessary to perform scale normalization on both metrics separately to eliminate the influence of dimensions and numerical range.

[0050] After standardization, the standardized sensitivity metric and the standardized dynamic range metric are fused to obtain the raw importance score for each expert unit. This score already integrates information from two dimensions. However, for ease of subsequent comparison and use, a final processing step is required: normalizing the raw importance scores of all expert units, for example, scaling them to the [0, 1] interval, thereby generating the final importance score for each expert unit.

[0051] The technical advantage of this scheme lies in ensuring a more robust and fair calculation process for importance scores by introducing standardization and normalization steps. Scale standardization avoids fusion bias caused by numerical differences, while the final normalization allows for a direct comparison of the importance of different expert units and provides convenient input for subsequent score-based bit-width allocation algorithms.

[0052] In a more specific implementation, the aforementioned scale standardization and fusion calculation process can be achieved using logarithmic transformation and weighted linear fusion. Specifically, for sensitivity measurement... and dynamic range measurement Logarithmic transformations are performed separately to compress the numerical range and reduce the impact of extreme values. The standardized sensitivity metric after transformation is: The standardized dynamic range metric is The calculation formula is as follows:

[0053]

[0054] in, It is the natural logarithm function. For a preset very small positive number (e.g.) Its function is to prevent when or A calculation error occurs when the value is zero.

[0055] Obtain the standardized metric and Subsequently, a weighted linear fusion method was used to calculate the original importance scores of the expert units. The calculation formula is as follows:

[0056] in, and These are all preset fusion weight coefficients, which can be set based on experience or experimentation to adjust the relative contributions of sensitivity and dynamic range in the final importance assessment. For example, you can set... To give both equal weight.

[0057] The technical advantage of this scheme lies in providing a more refined and flexible method for calculating importance scores. Logarithmic transformation is an effective technique for handling data with long-tailed distributions, making importance scores less sensitive to drastic fluctuations in numerical values. The weighted linear fusion approach provides an adjustable mechanism, allowing developers to balance the importance of parameter sensitivity and the dynamic range of information according to the specific needs of the task, thereby fine-tuning the quantification strategy.

[0058] In another preferred embodiment, the steps for bit width allocation based on importance scores are described in detail. This process aims to convert importance scores into specific bit width values ​​while satisfying given resource constraints. First, based on the final importance score of each expert unit... (Assuming it has been normalized to the [0, 1] interval), it is initially converted into a continuous bit width requirement value through linear mapping. The calculation formula is as follows:

[0059] in, and These are the preset upper and lower bounds of the bit width requirement, for example, 8 and 4 respectively. In this way, the most important unit will be mapped to... The lowest will be mapped to .

[0060] Then, it is necessary to check the current bit width allocation scheme (by all...) Does the current resource configuration meet the preset resource constraints? Therefore, it is necessary to calculate the total resource consumption under the current scheme. When the resource constraint is the model storage size, the total resource consumption can be calculated as the sum of the products of the number of parameters and the bit width of all expert units: ,in It is the first The number of parameters to be quantized in each expert unit. When the constraint is inference latency, the total resource consumption can be determined by a lookup table function. To calculate: The lookup table is pre-calibrated based on the target hardware platform and provides the inference latency corresponding to different bit widths.

[0061] If calculated Exceeding the preset resource budget This necessitates adjusting the bit width allocation. This can be structured as an optimization problem: under constraints... Below, maximize the total weighted importance score. .in, It is a resource consumption function. It is a monotonically non-decreasing function of bit width, representing the benefit brought by bit width. By solving this optimization problem (e.g., using iterative adjustment, dynamic programming, or heuristic search), a set of continuous bit width requirements can be obtained that satisfies resource constraints and optimizes the overall performance of the model. .

[0062] Finally, since hardware typically only supports discrete bit width values ​​(e.g., {4, 6, 8}), the optimized continuous bit width requirement needs to be considered. Projected onto candidate bit-width set In this process, the final discrete target quantization bit width of each expert unit is obtained. A simple and effective method is to project to the nearest set, i.e., select a set. Zhongyu Closest value:

[0063] The technical advantage of this approach lies in its formalization of the bit-width allocation problem into a constrained optimization problem, enabling a systematic search for the optimal bit-width configuration within hardware resource limitations. This is far more accurate than simple linear mapping followed by truncation, ensuring that resources are utilized most efficiently in exchange for maximum model performance preservation. This method is highly versatile and adaptable to different types of resource constraints (storage, latency, etc.) and hardware platforms.

[0064] Furthermore, a discrete target quantization bit width was allocated to each expert unit. The specific steps for generating the initial hybrid bit-width model are as follows. First, for each expert unit... It needs to be based on its allocated bit width This determines the range of values ​​for the quantized integer representation. For signed integers, this range is... The minimum value maximum value For example, for a 4-bit signed integer, the range is [-8, 7].

[0065] Then, for this expert unit weight or activation value Construct a pseudo quantization operator This operator simulates the quantization process from floating-point to fixed-point, and its specific form is as follows:

[0066] in, It is the quantization step size. These are quantization zeros, which are learnable parameters used to quantize floating-point values. Mapped to the integer field. The function performs rounding operations (such as rounding to the nearest integer). The function then truncates the rounded result to the previously defined range of integers. In the actual forward propagation, this integer value also needs to be dequantized back to the floating-point domain (by multiplying by...). and subtract ), to continue the subsequent calculations of the network.

[0067] Finally, a configuration will be built for each expert unit, along with its bit width. The corresponding pseudo-quantization operator is inserted before the weights and / or activation computation paths of the corresponding expert units in the original floating-point neural network model, replacing the original floating-point operations. After the operator insertion of all expert units is completed, the initial mixed-precision quantization model is obtained. This model is structurally similar to the floating-point model, but it simulates the effect of mixed-precision quantization during the computation process.

[0068] The technical advantage of this scheme lies in its clear definition of how to implement mixed-bit-width pseudo-quantization in neural networks. By introducing numerical ranges corresponding to the bit widths of each expert unit and learnable quantization parameters (step size and zeros), the model can adaptively learn the optimal quantization scheme for each layer during training, thereby minimizing quantization error. This is the foundation for achieving high-performance quantization-aware fine-tuning.

[0069] In the quantization-aware adjustment stage of the initial hybrid bit-width model, a preferred implementation involves a specially designed loss function. Specifically, a weighted total loss function is constructed for quantization-aware adjustment. Its form is as follows:

[0070] in, This is the total number of expert units. It is calculated based on the task dataset and is related to the first... Each expert unit outputs the relevant original task loss components. The key to this invention is the expert-perceived weight, which is related to the first... The final importance score of each expert unit Positive correlation. It is an optional quantization regularization term, and It is its corresponding regularization coefficient.

[0071] To ensure that expert units with higher importance have a larger weight in the loss function, expert-perceived weights are used. It can be based on the final importance score of each expert unit. To set it. One effective way is to use a system with temperature hyperparameters. The calculation is performed using the Softmax function:

[0072] in, It is a positive number used to adjust the smoothness of the weight distribution. When When the weight is smaller, the weight will be more concentrated on the few expert units with the highest importance; when When the weights are larger, the weight distribution becomes more gradual. In this way, the importance score... It was transformed into weights to guide the training process. .

[0073] After constructing the weighted loss function Then, using the task dataset and this loss function as the optimization objective, the initial mixed bit-width model is iteratively trained multiple times (e.g., via gradient descent). During training, because the loss components of important units are amplified, the optimization algorithm prioritizes adjusting parameters to reduce the errors of these units. The training process continues until a preset stopping condition is met (e.g., reaching the maximum number of training rounds or the validation set performance no longer improving). The model obtained at this point is the final, fully optimized mixed bit-width quantization model.

[0074] By introducing a weighted loss function linked to expert importance scores, "expert-aware" fine-tuning is achieved. This allows training resources to be allocated more intelligently, prioritizing the performance of key expert units, thereby better maintaining the overall model accuracy and the reliability of risk assessment. Compared to the traditional QAT method that treats all units equally, this approach achieves better post-quantization performance with less fine-tuning, and is particularly suitable for non-homogeneous network structures containing a few key modules.

[0075] Furthermore, in the aforementioned weighted loss function, the quantization regularization term... Various forms can be adopted to constrain the quantization process from different perspectives, reducing the difference between the floating-point model and the quantization model. This regularization term... It can be a combination of one or more of the following: weight quantization error regularization term. Activate the distribution alignment regularization term and gradient preservation regularization term .

[0076] Weight quantization error regularization term The difference in weights before and after quantization is directly penalized, and it is defined as follows:

[0077] in, It is the first The original floating-point weights of each expert unit, These are the weights after applying the pseudo-quantization operator. This represents the square of the L2 norm. This term contributes to the learnable quantization parameters. Adjust it to a position that makes the quantized weights closest to the original floating-point weights.

[0078] Activate distribution alignment regularization term The aim is to maintain the consistency of the activation value distribution before and after quantization. It is defined as:

[0079] in, It is the first The set of floating-point activation values ​​of each expert unit on the task dataset. It is the set of quantized activation values. This indicates obtaining its probability distribution, while This is the Kullback-Leibler divergence, used to measure the difference between two probability distributions. This term helps maintain the information flow between network layers and prevents distribution drift caused by quantization.

[0080] Gradient Preservation Regularization This focuses on maintaining the learning dynamics of the model. It is defined as:

[0081] in, In a floating-point model, the task loss is relative to the weights. The gradient, and This refers to the gradient in the quantization model. This project encourages that the gradient direction and magnitude of the quantization model be consistent with those of the floating-point model, which helps to ensure stable convergence of quantization-aware training.

[0082] The technical advantage of this approach lies in providing richer and more direct supervisory signals for quantization-aware fine-tuning by introducing these structured regularization terms. These terms constrain the behavior of the quantization model from three levels: numerical, distributional, and gradient, enabling it to more comprehensively mimic the original floating-point model. This results in a final quantization model with higher performance than fine-tuning using only task loss.

[0083] In a specific application scenario, the method described in this invention is applied to a risk perception network for predicting the remaining useful life (RUL) of an aero-engine. In this scenario, the floating-point neural network model has a specific structure, typically comprising a shared feature encoder and at least one risk quantization output head. The shared feature encoder is responsible for extracting high-dimensional features from the raw sensor time-series data, while the risk quantization output head is designed as an independent expert unit whose task is to output conditional quantile estimates corresponding to a preset risk level based on the extracted features. For example, one output head might predict the 5th quantile of the RUL, another the 50th quantile (i.e., the median), and yet another the 95th quantile, thus forming a complete characterization of future lifetime uncertainty.

[0084] In this application, the risk quantization output heads are the most critical part of the network because they are directly related to the final risk assessment decision. Therefore, when assigning discrete target quantization bit widths to expert units, an additional protective rule can be introduced: the target quantization bit widths assigned to these risk quantization output heads are forcibly set to be no less than a preset bit width protection threshold. For example, regardless of their calculated importance score, these output heads are ensured to be allocated at least 8 bits or more, while the bit widths of other parts of the network can fluctuate between 4 and 8 bits depending on their importance score.

[0085] The technical advantage of this solution lies in its close integration of general quantization methods with the specific needs of high-risk application scenarios. By implementing mandatory high-bit-width protection for key output heads that directly impact risk decisions, the possibility of serious deviations in risk assessment due to quantization errors can be significantly reduced. This provides an important safety guarantee for deploying deep learning models in the safety-critical aviation field, ensuring that the model's core risk perception capabilities are not compromised while it is compressed and accelerated.

[0086] In this embodiment, the neural network model processed is a risk perception network specifically designed for RUL prediction. This network includes a shared feature encoder consisting of convolutional and recurrent layers, and three independent output heads. These three output heads, acting as independent expert units, are used to predict the 5th, 50th, and 95th percentiles of the RUL, respectively, thereby quantifying the engine failure risk. The goal of quantization is to maximize the model's prediction accuracy within the constraint that the total model size does not exceed 10MB.

[0087] First, a pre-trained floating-point RUL prediction model and a calibration dataset containing engine sensor data under various operating conditions are obtained. Then, the system performs an expert profiling process. Through one forward and backpropagation pass on the calibration dataset, the sensitivity metric of each convolutional layer, recurrent layer, and three output heads (i.e., all expert units) in the network is calculated. and dynamic range measurement The calculation formula follows the aforementioned embodiment, namely... The expected value of the squared gradient. The variance of the activation values.

[0088] Next, importance fusion is performed. The system uses logarithmic transformation to... and Standardize the scale, that is, calculate and Subsequently, the original importance score was calculated using equal-weighted linear fusion. Finally, for all expert units... Min-max normalization is performed to obtain the final importance score in the range [0, 1]. .

[0089] Then, the dynamic bit width allocation phase begins. The candidate bit width set is set as follows: Bits. The system first determines the importance score. Map each unit to a continuous bit-width requirement value. Next, it constructs and solves an optimization problem with the goal of minimizing the total model size. Under constraints, maximize the overall performance of the model (by... (Approximate). During the solution process, a specific rule was added: for the three risk quantization output heads, their allocated bit width must be 8 bits. After solving the optimization problem, the optimal continuous bit width for each unit was obtained. Then, by projection onto the nearest neighbor, the final discrete target quantization bit width is obtained. .

[0090] Subsequently, the system determines the bit width based on the allocated bit width. Construct a pseudo-quantization operator for each expert unit. The numerical range according to Determine the quantization step size. and zero point These were set as learnable parameters. After inserting these operators into the network, an initial mixed bit-width model was generated.

[0091] Finally, quantitative fine-tuning of expert perception is performed. The total loss function used is... Among them, expert perception weight Importance score By having temperature The Softmax function is used for generation. The regularization term selects the weighted quantization error term. Alignment terms with activation distribution A combination of methods was used to constrain the model from both numerical and distributional perspectives. The model was fine-tuned for 20 epochs using the task dataset until the loss converged on the validation set.

[0092] The complete workflow of this embodiment is as follows: Input a pre-trained floating-point RUL model and a hardware resource budget. First, automatically perform expert profiling, calculate Fisher information and activation variance for all layers, and fuse them into a normalized importance score. Second, solve a constrained optimization problem, automatically assigning a 4-bit or 8-bit bit width to each layer while ensuring all output headers are 8 bits. Third, automatically insert corresponding spurious quantization nodes into the model according to the bit width allocation scheme. Fourth, use a loss function weighted by importance scores and containing multiple regularization terms to fine-tune the model over short periods. Fifth, output a hybrid bit-width quantization model that meets resource budget requirements, performance optimization, and is ready for direct deployment.

[0093] By combining all the aforementioned technical features, this embodiment achieves significant comprehensive technical effects. Firstly, in terms of performance, due to the mandatory 8-bit width protection for the three risk quantization output heads, combined with expert-perceived weighted fine-tuning, the final quantization model shows minimal difference from the original floating-point model in key indicators of RUL prediction (such as RMSE and quantile loss), ensuring the reliability of risk assessment. Secondly, in terms of efficiency, by allocating 4-bit widths to a large number of low-importance convolutional and recurrent layers, the total model size is successfully controlled to within 10MB, achieving significant compression compared to the original floating-point model (typically exceeding 50MB) and the unified 8-bit quantization model (approximately 25MB), significantly reducing storage and loading overhead on avionics equipment. Finally, in terms of automation, the entire quantization process, from importance assessment to fine-tuning, is data-driven and automatically executed, eliminating the need for tedious manual parameter tuning and rule setting, greatly improving R&D efficiency. Therefore, the synergistic effect of these technical features enables this invention to achieve efficient and reliable deployment of neural network models while meeting stringent aerospace application requirements.

[0094] The technical solutions provided in this invention are not limited to the health management of aero-engines. Their core idea, balancing performance and resource consumption through expert-perceived hybrid bit-width quantization, has broad applicability. Any deep learning application that needs to be deployed on resource-constrained edge devices and includes multiple modules with varying degrees of impact on performance can benefit from this invention.

[0095] A typical application area is predictive maintenance in the Industrial Internet of Things (IIoT). For example, quantized neural network models can be deployed on critical industrial equipment such as large wind turbines, pump sets in chemical plants, or CNC machine tools to monitor equipment status in real time, predict faults, and determine remaining lifespan. In these scenarios, model inference must be performed on edge controllers close to the equipment, with strict requirements on latency and power consumption. Furthermore, the output layer responsible for final fault diagnosis or lifespan prediction is far more important than the front-end feature extraction layer. This invention allows for the allocation of high bit widths to these critical output layers, ensuring decision accuracy, while simultaneously performing deep compression on other layers to meet hardware limitations.

[0096] Another important application area is autonomous driving. Perception models in autonomous driving systems, such as CNNs for processing camera images or models for processing LiDAR point clouds, are typically large-scale and computationally intensive. Model quantization is an essential step for achieving real-time processing on in-vehicle computing platforms (such as NVIDIA Drive Orin and Qualcomm Snapdragon Ride). Modules responsible for detecting small targets (such as distant pedestrians) or identifying key traffic signs in these models are extremely sensitive; any tiny quantization error can lead to catastrophic consequences. The method of this invention can automatically identify these "expert units" and protect them with high bit width, thereby ensuring the safety of the autonomous driving system while achieving overall acceleration.

[0097] In the field of medical devices, especially portable or wearable smart medical devices, this invention also has enormous application potential. For example, in smartwatches used for real-time electrocardiogram (ECG) anomaly detection, the built-in processor resources are very limited. Applying this invention to quantize the ECG analysis model allows it to run for extended periods with low power consumption. The classification head in the model, responsible for identifying fatal arrhythmias (such as ventricular fibrillation), is a crucial expert unit. This invention ensures that it maintains a very high recall rate after quantization, avoiding missed fatal risks.

[0098] Furthermore, in fintech fields, such as real-time risk assessment in high-frequency trading, models need to run on dedicated FPGA or ASIC hardware with extremely low latency. The method proposed in this invention can guide the joint optimization of hardware-aware neural network architecture search (NAS) and quantitative strategies. By optimizing the hardware latency model (e.g., ... By introducing the optimization objective of bit width allocation, a hybrid bit width quantization model with optimal latency that is highly matched to a specific hardware architecture can be generated.

[0099] Based on the same inventive concept, this invention also provides a risk perception neural network quantification device for aero-engine health management, as described in the following embodiments. Since the principle of the risk perception neural network quantification device for aero-engine health management is similar to that of the risk perception neural network quantification method for aero-engine health management, the implementation of the risk perception neural network quantification device for aero-engine health management can refer to the implementation of the risk perception neural network quantification method for aero-engine health management, and repeated details will not be elaborated further. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0100] Figure 2 This is a structural block diagram of a risk perception neural network quantification device for aircraft engine health management according to an embodiment of the present invention, such as... Figure 2 As shown, it includes: importance calculation module 201, discrete bit width allocation module 202, and risk perception network construction module 203. The structure is described below.

[0101] Importance calculation module 201 is used to obtain a pre-trained floating-point neural network model and a representative dataset for calibration. For each expert unit in the floating-point neural network, based on the representative dataset, it calculates the sensitivity metric and dynamic range metric of the expert unit. It then fuses the sensitivity metric and dynamic range metric corresponding to each expert unit to generate the importance score of the expert unit. The discrete bit width allocation module 202 is used to allocate a discrete target quantization bit width to the expert unit based on the preset resource constraints and candidate bit width set, according to the importance score of the expert unit, wherein the level of the importance score allocation is proportional to the level of the discrete target quantization bit width. The risk-aware network construction module 203 is used to insert a pseudo-quantization operator corresponding to the target quantization bit width assigned to the expert unit in each expert unit of the floating-point neural network model, generate an initial mixed bit width model, perform quantization-aware adjustment on the initial mixed bit width model based on the task dataset, and introduce weights related to the importance score of the expert unit into the loss function of the initial mixed bit width model to generate a final mixed bit width quantization model, which serves as the risk-aware neural network.

[0102] In one embodiment, the importance calculation module includes: A forward propagation unit is used to perform a forward propagation process on the floating-point neural network model using the representative dataset; A dynamic range metric calculation unit is used to calculate, for each expert unit, the variance of the output activation values ​​on the representative dataset, and use it as the dynamic range metric for that expert unit. ,in, The output activation value of the expert unit; A backpropagation unit is used to perform a backpropagation process on the floating-point neural network model using the representative dataset. Sensitivity metric calculation unit, used to collect the first [value] for each expert unit. Trainable parameters of each expert unit The gradient value is calculated, and the expected value of the square of the gradient value on the representative dataset is used as a sensitivity measure of the expert unit. ,in, As expected, For partial differential equations, is the loss function of the floating-point neural network model.

[0103] In one embodiment, the importance calculation module further includes: Construct a floating-point neural network unit for the risk perception network of the floating-point neural network model for predicting remaining useful life, the floating-point neural network model including a shared feature encoder and at least one risk quantization output head; A preset risk level unit is used for the risk quantification output head to be an independent expert unit and is configured to output a conditional quantile estimate corresponding to the preset risk level. The target quantization bit width setting unit is used to set the target quantization bit width allocated to the risk quantization output head to a value not lower than a preset bit width protection threshold when allocating discrete target quantization bit width to the expert unit.

[0104] In one embodiment, the importance calculation module further includes: Importance scoring unit, used to assess the sensitivity measure for each expert unit. and the dynamic range metric The sensitivity metric and the dynamic range metric after standardization are fused and calculated to obtain the original importance score of the expert unit. The final importance score calculation unit is used to normalize the original importance scores of all expert units and generate the final importance score for each expert unit.

[0105] In one embodiment, the importance scoring unit is also used to measure the sensitivity. and the dynamic range metric Perform logarithmic transformations on each sample to generate a standardized sensitivity metric. and the dynamic range metric after standardization ,in, , ε is a preset minimum normal number; weighted linear fusion is used to calculate the original importance score of the expert unit. , where α and β are preset fusion weight coefficients.

[0106] In one embodiment, allocating a discrete bit-width module includes: The bit width requirement calculation unit is used to calculate the importance score of each expert unit. The expert unit Mapped to continuous bit width requirement values ,in, This is the upper bound of the preset bit width requirement. This is the lower bound of the preset bit width requirement; The total resource consumption calculation unit is used to calculate the total resource consumption under the current bit width allocation scheme. When the resource constraint is a storage size constraint, the total resource consumption under the current bit width allocation scheme is... , For the first The number of parameters to be quantized in each expert unit, and the total resource consumption under the current bit-width allocation scheme when the resource constraint is an inference delay constraint. , For lookup table functions based on the bit width to inference delay pre-calibrated for the target hardware platform; Importance score calculation unit, used under constraints Next, an optimization objective function is constructed, which is to maximize the total weighted importance score. ,in, The preset resource constraint threshold, This is a resource consumption function. It is a monotonically non-decreasing function with respect to bit width; The bit-width optimization unit is used to adjust the bit-width requirement value through the optimization objective function. To satisfy Furthermore, the bit width is optimized until a bit width requirement value that satisfies the aforementioned resource constraints is obtained. ; Discrete target quantization bit width calculation unit, used to meet the bit width requirement value of the resource constraints. Projected onto the candidate bit-width set In this process, the discrete target quantization bit width of each expert unit is obtained. .

[0107] In one embodiment, the risk perception network building module includes: Numerical range de-top cell, used for each expert cell Quantization bit width based on the allocated discrete target Determine the expert unit The range of integer values Among them, the minimum value maximum value ; Construct a pseudo-quantization operator unit for the expert unit. activation value Constructing a pseudo-quantization operator ,in, For learnable quantization step size parameters, To quantize the zero-point parameters, This is a truncation function. It is a rounding function; An initial mixed bit-width model construction unit is used to insert the pseudo-quantization operator before the weights and / or activation paths of each expert unit in the floating-point neural network model, replacing the original floating-point operations, to obtain the initial mixed bit-width model.

[0108] In one embodiment, the risk perception network building module further includes: Construct a loss function unit to build a weighted loss function for quantization-sensory adjustment. ,in, The total number of expert units. For calculation based on the task dataset and related to the first Each expert unit outputs the relevant original task loss components. In order to be with the first The final importance score of each expert unit Positively correlated expert perception weights The regularization coefficient is . To quantify the regularization term; Define perception weight units to be used based on the final importance score of each expert unit. Set the expert perception weight ,in, This refers to temperature hyperparameters. A model generation unit is used to generate models based on the task dataset and the expert-perceived weights. The constructed weighted loss function To achieve this, the initial mixed bit-width model is trained iteratively for multiple rounds until a preset stopping condition is met, thereby obtaining the final mixed bit-width quantization model.

[0109] In one embodiment, the loss function unit is also used for the quantization regularization term. For weight quantization error regularization term Activate the distribution alignment regularization term Gradient Preservation Regularization Term One or more combinations thereof; weight quantization error regularization term ,in, For the first The original floating-point weights of each expert unit, For pseudo-quantization operators, For learnable quantization step size parameters, To quantize the zero-point parameters; activate the distribution alignment regularization term. ,in, For the first The set of activation values ​​for each expert unit on the task dataset. For probability distribution, Kullback-Leibler divergence; gradient-preserving regularization term ,in, For the first The gradient of the task loss for each expert unit in the floating-point model. This represents the gradient of the task loss in the quantization model.

[0110] The embodiments of the present invention achieve the following technical effects: In terms of performance, due to the mandatory 8-bit width protection of the three risk quantization output heads and the combination of expert-perceived weighted fine-tuning, the final quantization model shows minimal difference from the original floating-point model in key indicators of RUL prediction (such as RMSE and quantile loss), ensuring the reliability of risk assessment. In terms of efficiency, by allocating 4-bit widths to a large number of low-importance convolutional and recurrent layers, the total model size is successfully controlled to within 10MB, achieving significant compression compared to the original floating-point model (typically exceeding 50MB) and the unified 8-bit quantization model (approximately 25MB), significantly reducing storage and loading overhead on avionics equipment. In terms of automation, the entire quantization process, from importance assessment to fine-tuning, is data-driven and automatically executed, eliminating the need for tedious manual parameter tuning and rule setting, greatly improving R&D efficiency. The synergistic effect of these technical features enables the embodiments of this invention to achieve efficient and reliable deployment of neural network models while meeting stringent aerospace application requirements.

[0111] Obviously, those skilled in the art should understand that the modules or steps of the above-described embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

[0112] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A risk perception neural network quantification method for aero-engine health management, characterized in that, include: Obtain a pre-trained floating-point neural network model and a representative dataset for calibration. For each expert unit in the floating-point neural network, calculate the sensitivity metric and dynamic range metric of the expert unit based on the representative dataset. Fuse the sensitivity metric and dynamic range metric corresponding to each expert unit to generate the importance score of the expert unit. Based on preset resource constraints and a candidate bit width set, a discrete target quantization bit width is allocated to the expert unit according to the importance score of the expert unit, wherein the level of the importance score is proportional to the level of the discrete target quantization bit width. In each expert unit of the floating-point neural network model, a pseudo-quantization operator corresponding to the target quantization bit width assigned to the expert unit is inserted to generate an initial mixed bit width model. The initial mixed bit width model is then subjected to quantization-aware adjustment based on the task dataset. Furthermore, weights related to the importance score of the expert unit are introduced into the loss function of the initial mixed bit width model to generate the final mixed bit width quantization model, which serves as the risk-aware neural network.

2. The risk perception neural network quantification method for aero-engine health management as described in claim 1, characterized in that, For each expert unit in the floating-point neural network, based on the representative dataset, the sensitivity metric and dynamic range metric of the expert unit are calculated, including: Using the representative dataset, perform a forward propagation process on the floating-point neural network model; For each expert unit, the variance of the output activation values ​​on the representative dataset is calculated and used as a measure of the dynamic range of the expert unit. ,in, The output activation value of the expert unit; Using the representative dataset, perform a backpropagation process on the floating-point neural network model; For each expert unit, collect the first... Trainable parameters of each expert unit The gradient value is calculated, and the expected value of the square of the gradient value on the representative dataset is used as a sensitivity measure of the expert unit. ,in, As expected, For partial differential equations, is the loss function of the floating-point neural network model.

3. The risk perception neural network quantification method for aero-engine health management as described in claim 1, characterized in that, The sensitivity metric and dynamic range metric corresponding to each expert unit are fused to generate the importance score of the expert unit, including: The sensitivity measure for each expert unit and the dynamic range metric The sensitivity metric and the dynamic range metric after standardization are fused together to obtain the original importance score of the expert unit. The original importance scores of all expert units are normalized to generate the final importance score for each expert unit.

4. The risk perception neural network quantification method for aero-engine health management as described in claim 3, characterized in that, The sensitivity measure for each expert unit and the dynamic range metric Scale standardization is performed separately, and the standardized sensitivity metric and the standardized dynamic range metric are fused together to obtain the original importance score of the expert unit, including: For the sensitivity measure and the dynamic range metric Perform logarithmic transformations on each sample to generate a standardized sensitivity metric. and the dynamic range metric after standardization ,in, , ε is a preset minimum positive constant; The original importance score of the expert unit is calculated using weighted linear fusion. , where α and β are preset fusion weight coefficients.

5. The risk perception neural network quantification method for aero-engine health management as described in claim 1, characterized in that, Based on preset resource constraints and a candidate bit width set, and according to the importance score of the expert unit, a discrete target quantization bit width is allocated to the expert unit, including: Based on the importance score of each expert unit The expert unit Mapped to continuous bit width requirement values ,in, This is the upper bound of the preset bit width requirement. This is the lower bound of the preset bit width requirement; Calculate the total resource consumption under the current bit width allocation scheme. When the resource constraint is a storage size constraint, the total resource consumption under the current bit width allocation scheme is... , For the first The number of parameters to be quantized in each expert unit, and the total resource consumption under the current bit-width allocation scheme when the resource constraint is an inference delay constraint. , For lookup table functions based on the bit width to inference delay pre-calibrated for the target hardware platform; Under constraints Next, an optimization objective function is constructed, which is to maximize the total weighted importance score. ,in, The preset resource constraint threshold, This is a resource consumption function. It is a monotonically non-decreasing function with respect to bit width; The bit width requirement is adjusted using the optimization objective function. To satisfy Furthermore, the bit width is optimized until a bit width requirement value that satisfies the aforementioned resource constraints is obtained. ; Bit width requirement to satisfy the resource constraints Projected onto the candidate bit-width set In this process, the discrete target quantization bit width of each expert unit is obtained. .

6. The risk perception neural network quantification method for aero-engine health management as described in claim 1, characterized in that, In each expert unit of the floating-point neural network model, a pseudo-quantization operator corresponding to the target quantization bit width assigned to the expert unit is inserted to generate an initial mixed bit width model, including: For each expert unit Quantization bit width based on the allocated discrete target Determine the expert unit The range of integer values Among them, the minimum value maximum value ; For the expert unit activation value Constructing a pseudo-quantization operator ,in, For learnable quantization step size parameters, To quantize the zero-point parameters, This is a truncation function. It is a rounding function; The pseudo-quantization operator is inserted before the weights and / or activation paths of each expert unit in the floating-point neural network model to replace the original floating-point operations, thus obtaining the initial mixed bit-width model.

7. The risk perception neural network quantification method for aero-engine health management as described in claim 1, characterized in that, The initial mixed bit-width model is quantized and adjusted based on the task dataset, and weights related to the importance scores of the expert units are introduced into the loss function of the initial mixed bit-width model to generate the final mixed bit-width quantization model, including: Construct a weighted loss function for quantizing perception adjustment. ,in, The total number of expert units. For calculation based on the task dataset and related to the first Each expert unit outputs the relevant original task loss components. In order to be with the first The final importance score of each expert unit Positively correlated expert perception weights The regularization coefficient is . To quantize the regularization term; Based on the final importance score of each expert unit Set the expert perception weight ,in, This refers to temperature hyperparameters. Using the task dataset, based on the expert-perceived weights The constructed weighted loss function To achieve this, the initial mixed bit-width model is trained iteratively for multiple rounds until a preset stopping condition is met, thereby obtaining the final mixed bit-width quantization model.

8. The risk perception neural network quantification method for aero-engine health management as described in claim 7, characterized in that, The quantization regularization term For weight quantization error regularization term Activate the distribution alignment regularization term Gradient preservation regularization term One or more combinations thereof; Weight quantization error regularization term ,in, For the first The original floating-point weights of each expert unit, For pseudo-quantization operators, For learnable quantization step size parameters, To quantize the zero-point parameters; Activate distribution alignment regularization term ,in, For the first The set of activation values ​​for each expert unit on the task dataset. For probability distribution, The Kullback-Leibler divergence; Gradient Preservation Regularization ,in, For the first The gradient of the task loss for each expert unit in the floating-point model. This represents the gradient of the task loss in the quantization model.

9. The risk perception neural network quantification method for aero-engine health management as described in any one of claims 1 to 8, characterized in that, The floating-point neural network model is used as a risk perception network for predicting remaining useful life. The floating-point neural network model includes a shared feature encoder and at least one risk quantization output head. The risk quantification output head is an independent expert unit and is configured to output a conditional quantile estimate corresponding to a preset risk level. When allocating discrete target quantization bit width to the expert unit, the target quantization bit width allocated to the risk quantization output head is set to a value not lower than a preset bit width protection threshold.

10. A risk perception neural network quantification device for aircraft engine health management, characterized in that, include: The importance calculation module is used to obtain a pre-trained floating-point neural network model and a representative dataset for calibration. For each expert unit in the floating-point neural network, based on the representative dataset, the module calculates the sensitivity metric and dynamic range metric of the expert unit. The module then fuses the sensitivity metric and dynamic range metric corresponding to each expert unit to generate the importance score of the expert unit. The discrete target quantization bit width allocation module is used to allocate discrete target quantization bit width to the expert unit based on preset resource constraints and a candidate bit width set, according to the importance score of the expert unit, wherein the level of the allocated importance score is proportional to the level of the discrete target quantization bit width. A risk-aware neural network construction module is used to insert a pseudo-quantization operator corresponding to the target quantization bit width assigned to the expert unit in each expert unit of the floating-point neural network model to generate an initial mixed bit width model, perform quantization-aware adjustment on the initial mixed bit width model based on the task dataset, and introduce weights related to the importance score of the expert unit into the loss function of the initial mixed bit width model to generate a final mixed bit width quantization model, which serves as the risk-aware neural network.

Images