Heterogeneous sensor dynamic compensation method and device based on large model knowledge prior and conditional parameter generation

Abstract

The application discloses a heterogeneous sensor dynamic compensation method and device based on large model knowledge prior and conditional parameter generation, and relates to the technical field of measurement. The method comprises the following steps: acquiring multi-source data of a target sensor, including meta information, a knowledge prior vector and a sample difference descriptor; mapping the multi-source data into a conditional parameter of characteristic affine modulation; acquiring a dynamic response signal sequence of the target sensor, performing coarse compensation on the dynamic response signal sequence based on a physical model, and acquiring a physical coarse compensation sequence; fusing the dynamic response signal sequence and the physical coarse compensation sequence, constructing a joint input feature matrix, inputting the joint input feature matrix into a deep neural network of the conditional parameter, and acquiring a nonlinear residual correction signal; and acquiring a dynamic compensation output signal according to the physical coarse compensation sequence and the nonlinear residual correction signal. The application can improve the environmental adaptability of sensor compensation, so as to guarantee the measurement accuracy in variable working condition scenes.

Description

Technical Field

[0001] This application relates to the field of measurement technology, and in particular to a method and apparatus for dynamic compensation of heterogeneous sensors based on large model knowledge priors and conditional parameter generation. Background Technology

[0002] With the increasing demand for high-dynamic and high-precision testing in fields such as aerospace, weapons testing, intelligent manufacturing, and industrial process monitoring, the dynamic response capability of sensors under complex transient conditions has become a core constraint affecting the reliability of test results. In scenarios involving shock, vibration, sudden pressure changes, and multi-physics coupling, the measured signal typically exhibits characteristics such as suddenness, strong non-stationarity, and wide bandwidth. This not only requires the measurement system to have extremely fast response speed and a wide frequency response range, but also high dynamic consistency. However, in actual measurement engineering, sensors and their measurement links are inevitably affected by a combination of complex physical factors such as their limited bandwidth, time constant, damping characteristics, installation stiffness, and temperature drift. This results in significant dynamic errors in the actual output signal compared to the ideal reference signal, including time delay, amplitude attenuation, phase distortion, and resonance peak deviation. These errors severely weaken the sensor's ability to reproduce real physical events, directly leading to the failure of subsequent state recognition, feature extraction, and control decisions.

[0003] To overcome the aforementioned dynamic response distortion, current engineering practices typically employ sensor dynamic compensation technology to perform algorithmic correction and recovery of measured signals. However, in real-world, complex field testing environments, existing dynamic compensation systems generally face the serious challenge of "cross-condition failure." Taking aero-engine ground testing or structural impact testing as an example, a dynamic compensation model calibrated for a specific sensor under standard laboratory conditions often exhibits extremely high waveform recovery accuracy. However, when the sensor is deployed to a real field, drastic changes in ambient temperature or severe mechanical vibration can cause slight loosening of the mounting base, fundamentally altering the sensor's physical boundary conditions. Consequently, its natural resonant frequency and damping ratio drift significantly. In this case, the dynamic compensation model based on fixed parameters not only fails to effectively correct the distorted waveform but also easily amplifies high-frequency noise, leading to the complete failure of critical transient physical peak and arrival time assessments. This necessitates significant time and economic costs to re-acquire data and recalibrate the model for the new conditions.

[0004] Currently, research on sensor dynamic compensation mainly focuses on two types of methods: traditional physical low-order models and purely data-driven models. Traditional physical methods achieve dynamic recovery through first- or second-order inertial elements, system identification, and inverse filtering. These methods have clear physical meaning, but their compensation parameters are usually statically fixed, and they severely lack generalization ability when faced with multi-source heterogeneous differences caused by changes in temperature or installation stiffness. With the development of deep learning, data-driven models such as one-dimensional convolutional neural networks and temporal convolutional networks are widely used, and they have a strong fitting ability for nonlinear distortion. However, these end-to-end black-box models lack clear physical boundary constraints, and they must rely on massive amounts of new data for retraining when faced with changes in operating conditions, which is not conducive to rapid iteration and lightweight deployment under multiple sensor types.

[0005] Therefore, when faced with real-world engineering applications involving the sharing of heterogeneous sensors and frequent switching of complex operating conditions, existing methods cannot directly establish a collaborative mechanism between environmental perception and physical adaptation, making it difficult to overcome the degradation of model generalization performance and the high costs of recalibration and retraining caused by changes in operating conditions. Summary of the Invention

[0006] Therefore, it is necessary to provide a method and device for dynamic compensation of heterogeneous sensors based on the prior knowledge and condition parameters of a large model to address the above-mentioned technical problems. This method and device can effectively improve the model's adaptive and unified compensation capabilities for heterogeneous sensors under varying operating conditions.

[0007] Firstly, this application provides a method for dynamic compensation of heterogeneous sensors based on prior knowledge of a large model and the generation of conditional parameters. This method includes:

[0008] Acquire multi-source data from the target sensor, including metadata, prior knowledge vectors, and sample difference descriptors; among which, the prior knowledge vectors are extracted from the engineering text of the target sensor, and the sample difference descriptors are used to characterize the temporal error of the target sensor.

[0009] Mapping multi-source data to conditional parameters of affine modulation;

[0010] The dynamic response signal sequence of the target sensor is obtained, and coarse compensation is performed on the dynamic response signal sequence based on the physical model to obtain the physical coarse compensation sequence.

[0011] The dynamic response signal sequence and the physical coarse compensation sequence are fused to construct a joint input feature matrix. The joint input feature matrix is ​​then input into a deep neural network with conditional parameters to obtain the nonlinear residual correction signal.

[0012] The dynamic compensation output signal of the target sensor is obtained based on the physical coarse compensation sequence and the nonlinear residual correction signal.

[0013] In one embodiment, obtaining the prior knowledge vector includes:

[0014] A large language model is used to map unstructured engineering text into knowledge prior vectors.

[0015] In one embodiment, the conditional parameters for mapping multi-source data to feature affine modulation include:

[0016] Meta-information, knowledge prior vector, and sample difference descriptor are concatenated into a composite prior representation vector, which is then input into a multilayer perceptron to obtain conditional parameters.

[0017] In one embodiment, coarse compensation is performed on the dynamic response signal sequence based on a physical model to obtain a physical coarse compensation sequence, including:

[0018] Based on meta-information and prior knowledge vectors, a physical inverse model of the target sensor is established;

[0019] Based on the physical inverse model, a regularization penalty factor is introduced in the frequency domain to construct a regularized inverse operator;

[0020] The inverse regularization operator is converted into a time-domain FIR filter system, and the dynamic response signal is coarsely compensated by the FIR filter system to obtain a physical coarse compensation sequence.

[0021] In one embodiment, the deep neural network includes a plurality of stacked residual blocks, each residual block including a one-dimensional causal dilated convolutional layer;

[0022] The joint input feature matrix is ​​input into a deep neural network to obtain the nonlinear residual correction signal, including:

[0023] After each residual block completes the convolution operation, the conditional parameters are loaded to perform a channel-level feature affine transformation.

[0024] In one embodiment, few-shot transfer learning is performed in response to changes in the target sensor type or operating conditions; wherein, few-shot transfer learning includes cutting off the updates of feature extraction convolution kernels shared across tasks in the physical model and deep neural network, and opening up the model weight updates used to generate conditional parameters.

[0025] Secondly, this application also provides a heterogeneous sensor dynamic compensation device based on large model knowledge priors and condition parameter generation. The device includes:

[0026] The offline prior configuration module is used to acquire multi-source data from the target sensor, including metadata, knowledge prior vectors, and sample difference descriptors. Among them, the knowledge prior vectors are extracted from the engineering text of the target sensor, and the sample difference descriptors are used to characterize the temporal error of the target sensor.

[0027] The conditional parameter generation module is used to map multi-source data into conditional parameters for feature affine modulation.

[0028] The online execution module is used to acquire the dynamic response signal sequence of the target sensor, perform coarse compensation on the dynamic response signal sequence based on the physical model to obtain the physical coarse compensation sequence, fuse the dynamic response signal sequence and the physical coarse compensation sequence to construct a joint input feature matrix, input the joint input feature matrix into a deep neural network with conditional parameters to obtain the nonlinear residual correction signal, and obtain the dynamic compensation output signal of the target sensor based on the physical coarse compensation sequence and the nonlinear residual correction signal.

[0029] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method for dynamic compensation of heterogeneous sensors based on large model knowledge priors and conditional parameters.

[0030] Fourthly, this application also provides a computer-readable storage medium. This computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps in the above-described method for dynamic compensation of heterogeneous sensors based on large model knowledge priors and conditional parameters.

[0031] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the steps in the above-described method for dynamic compensation of heterogeneous sensors based on large model knowledge priors and conditional parameters.

[0032] The aforementioned method and apparatus for dynamic compensation of heterogeneous sensors based on large-scale model knowledge priors and conditional parameters acquire multi-source data from the target sensor, including metadata, knowledge prior vectors, and sample difference descriptors. The knowledge prior vectors are extracted from the engineering text of the target sensor, and the sample difference descriptors characterize the temporal error of the target sensor. The multi-source data is mapped to conditional parameters for feature affine modulation. The dynamic response signal sequence of the target sensor is acquired, and coarse compensation is performed on the dynamic response signal sequence based on a physical model to obtain a physical coarse compensation sequence. The dynamic response signal sequence and the physical coarse compensation sequence are fused to construct a joint input feature matrix, which is then input into a deep neural network of conditional parameters to obtain a nonlinear residual correction signal. Based on the physical coarse compensation sequence and the nonlinear residual correction signal, the dynamic compensation output signal of the target sensor is obtained. This scheme effectively maps unstructured prior text knowledge into dynamic control conditions for the compensation network, reducing the burden of repetitive model training, improving cross-condition transferability, and thus meeting the extremely high requirements for high reliability of test data and lightweight model deployment in engineering applications. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating a heterogeneous sensor dynamic compensation method based on large model knowledge priors and condition parameters in one embodiment.

[0034] Figure 2 This is a schematic diagram of the multi-source data acquisition and condition parameter generation process in one embodiment;

[0035] Figure 3 This is a flowchart illustrating the physical coarse compensation and residual correction process in one embodiment.

[0036] Figure 4 This is a schematic diagram illustrating the joint optimization and migration deployment process of a heterogeneous sensor dynamic compensation system in one embodiment. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0038] To address the technical problems of existing computing frameworks failing to effectively utilize the external engineering text of heterogeneous sensors in complex transient tests, and the difficulty of generalizing existing models across operating conditions, this application provides a dynamic compensation method for heterogeneous sensors based on large model knowledge priors and conditional parameter generation. The overall technical approach is as follows: First, an offline prior configuration path is constructed, and unstructured equipment engineering text is mapped into knowledge prior vectors using a large language model; second, through a conditional parameter generation mechanism, the knowledge prior vectors are transformed into modulation parameters of the residual correction network; finally, in the online execution phase, the physical boundary of the system is established using a low-order physical model, and the nonlinear high-frequency error is corrected using the residual network controlled by the aforementioned conditional parameters.

[0039] Specifically, such as Figure 1 As shown, the method includes the following steps:

[0040] Step 101: Obtain multi-source data from the target sensor, including metadata, prior knowledge vectors, and sample difference descriptors; wherein, the prior knowledge vectors are extracted from the engineering text of the target sensor, and the sample difference descriptors are used to characterize the temporal error of the target sensor.

[0041] The processing objects of this invention not only include dynamic measured numerical sequences, but also encompass unstructured textual data that determines the dynamic characteristics of sensors. Specifically, for piezoresistive pressure sensors or strain gauge torque sensors exhibiting second-order oscillation characteristics, their prior information typically includes the inherent resonant frequency, damping ratio, and static calibration zero bias; for thermocouple sensors exhibiting first-order hysteresis, their prior information includes thermal response time and time constant. Traditional dynamic compensation networks, limited by their network architecture, can typically only receive a single one-dimensional time-series input and cannot directly calculate the aforementioned physical parameters described in natural language. To enable the compensation network to process and utilize textual prior information, this embodiment constructs a cross-modal data feature extraction and fusion mechanism, achieving a reliable mapping from unstructured text to continuous numerical vectors.

[0042] like Figure 2 As shown, for the heterogeneous target sensor to be compensated, two types of external information are acquired simultaneously: the first type is structured metadata. The first category includes sensor type, range, sampling rate, and operating temperature labels, which the system converts into an initial tensor through numerical normalization or encoding operations; the second category is unstructured engineering text. This includes equipment manuals, factory static calibration sheets, specification sheets, etc. To address the issues of semantic complexity and inconsistent formatting in the textual data, this embodiment employs a text embedding layer of a large language model to process the aforementioned text. The physical fields in the model undergo feature reduction and context encoding. During implementation, pre-defined engineering-extracted prompts guide the large language model to locate key physical parameters, and pooling features from the model's hidden states are extracted, outputting a fixed-dimensional sensor knowledge prior vector. :

[0043]

[0044] in, These are the basic weight parameters for large language models. This is an LLM editor function. The process transforms abstract textual properties into numerical representations that are differentiable within the computational framework.

[0045] Considering that the dynamic response of the target sensor in actual operation is also affected by specific test conditions, in order to introduce sample-level dynamic state information, this embodiment further extracts the time-domain error (such as rise time deviation, peak deviation ratio, main resonance peak offset, etc.) between the reference signal and the measured signal during the calibration stage or historical tests, and constructs a sample difference descriptor. In a zero-shot online testing scenario lacking any reference data, the descriptor is initialized with a zero-vector. Subsequently, the system concatenates static textual priors, metadata, and dynamic observation features through channel cascading operations to form a multi-dimensional composite prior representation vector. :

[0046]

[0047] This composite prior representation vector, within a unified feature space, encompasses both the physical performance benchmark of the target sensor and reflects the error drift trend under specific operating conditions, providing comprehensive input information for subsequent parameter generation.

[0048] Step 102: Map the multi-source data to the conditional parameters of the feature affine modulation.

[0049] In the multi-source data fusion stage, this invention uses a Feature-wise Linear Modulation (FiLM) affine transformation mechanism to directly apply the generated conditional parameters to the hidden layers within the residual correction network. Assuming the subsequent residual correction network contains several hidden layers, the conditional parameter generation module utilizes a Multilayer Perceptron (MLP) to receive the composite prior representation vector. and for the first in the network Each hidden layer generates specific conditional parameters, namely channel-level scaling factors used for feature affine modulation. With offset coefficient The generation and calculation process is represented as follows:

[0050]

[0051]

[0052] in, and These represent the learnable mapping weight matrix and bias term in the conditional parameter generation module, respectively. This structure enables the generative network to learn the nonlinear relationship between multimodal prior features and temporal compensation requirements.

[0053] During the forward propagation of the residual correction network, for the first... Intermediate feature maps output by the layer The system uses the above conditional parameters to perform channel-by-channel affine transform modulation:

[0054]

[0055] In the formula, This is an element-wise multiplication operation. Input the feature map before modulation. This is the output of the modulated feature map. In this mechanism, the scaling coefficients... The offset coefficient determines the degree of amplification or attenuation of the responses of different channels in the feature map. The trigger threshold of the channel activation function is then adjusted. Through the above formula, this invention substantially transforms external engineering knowledge into modulation parameters within the residual network. The network can then operate based on specific conditional parameters. and It adaptively adjusts the response weights of each channel in the feature map, thereby changing the feature extraction logic of the network while keeping the basic network topology fixed, and ultimately achieving effective model transfer across heterogeneous sensors and changing operating conditions.

[0056] Step 103: Obtain the dynamic response signal sequence of the target sensor, and perform coarse compensation on the dynamic response signal sequence based on the physical model to obtain the physical coarse compensation sequence.

[0057] In the dynamic compensation execution phase, this system constructs a collaborative compensation framework combining physical model coarse compensation and data-driven residual correction. Its core logic lies in decoupling the compensation task: utilizing a physical inverse model with a clear mathematical analytical form to recover the system's basic dynamic response profile, establishing the causality and stability boundaries of the signal; and then using a deep neural network to fit the unmodeled nonlinear residuals induced by complex variable operating conditions, solving the problem that traditional physical models struggle to cover high-frequency nonlinear distortions. The specific architecture and forward computation process of the collaborative network for physical coarse compensation and residual correction oriented towards dynamic boundary constraints are as follows: Figure 3 As shown.

[0058] For the input measured dynamic response signal sequence of the sensor to be compensated First, it is guided to the physical coarse compensation module. For engineering sensors widely used in engineering applications that exhibit first-order hysteresis or second-order oscillation characteristics, this module aims to initially recover the main amplitude-frequency attenuation and phase distortion caused by limitations in the sensor's inherent bandwidth. Let the physical model parameter set be... Its forward computation process is represented as follows:

[0059]

[0060] in, For physical coarse compensation sequence, This is a physical inverse operator, and a constrained finite-length unit impulse response (FIR) inverse filter. In practical implementation, the initialization parameter set is composed of sensor structured metadata extracted from multi-dimensional parameters and key physical parameters. The system establishes the basic order structure of the physical inverse model based on the sensor type label, substitutes the analyzed physical values ​​into it, and establishes the transfer function of the target sensor and discretizes it. .

[0061] To prevent high-frequency noise from being amplified infinitely during the direct inversion process, the system introduces a regularization penalty factor in the frequency domain. Construct the regularization inverse operator:

[0062]

[0063] in, for The conjugate transfer function.

[0064] The aforementioned operators are converted into time-domain FIR filter coefficients. In this step, the system does not rely on physical equations to approximate the real strong transient waveform containing complex coupled noise. Instead, it utilizes low-order dynamic mechanisms to establish a coarse compensation sequence that satisfies the system's causality and bounded-input-bounded-output (BIBO) stability. This mechanism removes the main linear delay component from the signal, significantly reducing the error mapping space that the subsequent neural network needs to fit, and fundamentally avoiding the non-causal lead response problem that easily arises in purely numerical end-to-end networks when dealing with high-frequency abrupt changes.

[0065] Step 104: The dynamic response signal sequence and the physical coarse compensation sequence are fused to construct a joint input feature matrix. The joint input feature matrix is ​​then input into a deep neural network with conditional parameters to obtain the nonlinear residual correction signal.

[0066] In obtaining the physical coarse compensation sequence Then, a deep neural network driven by conditional priors is activated; specifically, this deep neural network is a residual correction network. To achieve joint perception of initial distortion features and coarse compensation bias, the original measured dynamic response signal sequence and the physical model output are combined. Perform concatenated concatenation along the channel dimension to construct a joint input feature matrix. This operation preserves the high-frequency impact boundaries and nonlinear characteristics inherent in the original signal, providing ample clues for error evolution in the residual network.

[0067] At the network architecture level, the residual correction network in this embodiment adopts a one-dimensional temporal convolutional network (1D-TCN), whose high-frequency nonlinear residual correction signal... The mapping process is represented as:

[0068]

[0069] in, This represents a deep neural network with a feature modulation mechanism. This represents the learnable foundational weight matrix of the network. Specifically, the network consists of multiple stacked residual blocks, each containing a one-dimensional causal dilated convolution layer to ensure the causality of time-series processing and expand the receptive field. To achieve adaptive adjustment across operating conditions, the system synchronously loads conditional parameters (scaling coefficients) after the one-dimensional convolution operation and before the nonlinear activation function in each residual block. With offset coefficient This module performs channel-by-channel affine transformations. Its main functions are to suppress parasitic noise and distortion spurious peaks amplified by high-frequency gain extrapolation in the frequency domain, and to finely fit nonlinear distortions or transient hysteresis errors at extremely fast rising edges in the time domain.

[0070] Step 105: Obtain the dynamic compensation output signal of the target sensor based on the physical coarse compensation sequence and the nonlinear residual correction signal.

[0071] The system's dynamic compensation output signal From physical coarse compensation sequence With linear residual correction signal The linear addition obtained by performing time-domain synchronization is as follows:

[0072]

[0073] In this collaborative architecture, the physical inverse operator undertakes the fundamental task of establishing the basic dynamic response framework and system stability defense line; while the residual network, under the control of multimodal condition parameters, compensates for high-frequency nonlinear residuals based on sensor properties. This architecture effectively reduces the fitting difficulty and training cost of deep learning models, and achieves the complementary advantages of the rigor of physical laws and data-driven error correction under the condition of fixed network topology.

[0074] In one embodiment, after obtaining the physical coarse compensation results and condition parameters, the system enters the joint optimization and inference deployment phase. The core technical contribution of this phase lies in constructing a deep fusion mechanism between cross-modal features and time-series numerical computation. The specific implementation path encompasses three dimensions: joint training based on multi-objective physical constraints, asynchronous inference architecture with decoupled computing power, and lightweight parameter efficient fine-tuning (PEFT) for varying operating conditions. The workflow for system joint optimization and migration deployment is as follows: Figure 4 As shown.

[0075] During the model training phase, this system uses known high-precision reference signals. To supervise the benchmark, joint optimization is performed on the condition parameter generation module and the residual correction network. The joint input includes the measured dynamic sequence, the physical coarse compensation sequence, and prior knowledge and structured metadata parsed from the large language model. Traditional end-to-end networks typically rely solely on a single mean square error for loss evaluation, which can easily lead to over-smoothing or artifacts in the output waveform at local high-frequency features. To balance numerical fitting accuracy with the preservation of true dynamic features, this invention constructs a multi-objective joint loss function that includes time-domain, frequency-domain, and key dynamic indicators. :

[0076]

[0077] in, To balance the hyperparameter weights of each loss term.

[0078] Temporal reconstruction loss The global mean square error (MSE) is used to constrain the compensation network to approximate the true response in terms of overall signal energy and numerical amplitude.

[0079]

[0080] Frequency domain consistency loss The formula for calculating the amplitude-frequency attenuation and pseudo-resonance peak distortion of the compensation signal within the main operating frequency band is as follows:

[0081]

[0082] in This represents the Discrete Fourier Transform.

[0083] Key dynamic feature loss Regularization constraints are then applied to the specific physical properties of non-stationary transient signals. For example, in transient impact calibration experiments such as shock tubes, a differentiable feature extraction operator is constructed to directionally calculate the rise time of the predicted signal and the reference signal. ), maximum overshoot ( The response delay and the absolute error at the main resonant frequency are considered, and a penalty term is applied.

[0084] The aforementioned joint constraint mechanism ensures that while the modulated residual network approximates the low-frequency numerical solution, its high-frequency compensation output strictly conforms to the inherent physical response boundary and energy conservation law of the sensor.

[0085] During the engineering inference and actual deployment phases, the system adopts an asynchronous hardware and software collaborative architecture that decouples offline configuration from online compensation. This architecture effectively resolves the systemic contradiction between the large number of parameters in large language models and the long inference time, and the high sampling rate and low latency computation requirements of sensor transient testing.

[0086] Before performing a specific compensation task, the system opens an offline configuration channel: the system front-end interface receives unstructured text data (such as equipment manuals and static calibration sheets) from the sensor to be compensated, a large language model deployed on a high-computing-power node performs a one-time feature extraction on the text, and the condition parameter generation module forward calculates a fixed network condition parameter tensor (i.e., channel-level scaling factor). With offset coefficient Subsequently, the system serializes the static tensor and writes it to the storage area of ​​the edge inference device. This stage utilizes high-performance computing equipment to specifically process low-frequency, static, and complex semantic information.

[0087] During dynamic testing, the system switches to the online execution channel: at this time, the large language model module does not participate in the high-frequency calculation process at all, and the residual network (1D-TCN) at the edge is treated as a static constant and the above-mentioned fixed condition parameters are directly loaded. The sensor's megahertz-level high-frequency measured sequences are sequentially passed through a physical coarse compensation module and a modulation residual network, performing only pure numerical temporal convolution and feature multiplication-addition operations to output the final compensation result. This approach, while fully utilizing the text parsing capabilities of large language models, ensures extremely low memory usage and high real-time performance in the temporal signal processing from the underlying architecture.

[0088] In one embodiment, to address the data distribution shift and model applicability degradation caused by field sensor type change (such as cross-type replacement or new heterogeneous sensor access) or clamping condition drift (such as installation rigidity or ambient temperature), this invention implements a low-cost online adaptive migration strategy with a gradual transition based on a priori-driven mechanism.

[0089] First, when faced with new types of sensors or new test conditions that have not been trained on, the system performs zero-shot basic generalization based on knowledge reconstruction. The system only needs to input the unstructured text data or structured metadata of the new device into the offline configuration channel. Relying on a large language model and a conditional parameter generation module, the system regenerates the conditional parameters (i.e., scaling factors) for the new scenario. With offset coefficient In terms of network mechanism, the convolutional kernels of the underlying one-dimensional temporal convolutional network (1D-TCN) can capture fundamental dynamic features with physical universality (such as edge mutations, high-frequency decay trends, and other general temporal patterns); while the newly generated... and As a domain-specific controller, it directly updates the affine transformation gain weights within the residual network, thereby altering the activation distribution of the feature channels. This allows the network to realign the physical baseline based solely on textual priors, without ever accessing the measured data from this new type of sensor, achieving coarse-grained adaptive migration with zero samples.

[0090] If the current test scenario contains unknown strongly coupled noise or has extremely high accuracy requirements for dynamic errors, the system, based on zero-sample generalization, further collects a very small number of dynamic calibration signal sample pairs for this new type or new operating condition, and performs parameter-efficient fine-tuning (PEFT) on a small sample basis. During the specific backpropagation gradient calculation and weight update process, the system implements strict gradient freezing through a deep computation graph control strategy, forcibly cutting off the physical model parameter set. The main convolutional kernel weights responsible for general feature extraction in a one-dimensional temporal convolutional residual network The update path only allows gradient calculation and weight updates for the MLP parameter matrix within the condition parameter generation module. This method reduces the scale of parameter updates, which involves millions of parameters in traditional models, to an extremely small level. It significantly reduces the amount of calibration data, computing power consumption, and model retraining time required when introducing new heterogeneous sensors or facing changing operating conditions, enabling rapid adaptive deployment of heterogeneous sensor compensation models in complex engineering environments.

[0091] Traditional sensor compensation techniques have the following problems:

[0092] 1. Static parameter compensators lack generalization ability when faced with heterogeneous physical differences.

[0093] Existing traditional dynamic compensation methods are mostly based on low-order models, digital filters, or system identification. The structure and parameters of such models are usually calibrated for a single sensor, a specific range, or a fixed environment. However, when the sensor type, frequency response characteristics, clamping stiffness, or temperature environment in the engineering field drift, the physical boundary conditions change. The original fixed compensator parameters become unsuitable, requiring costly re-identification and physical calibration, and failing to form a unified compensation and online migration mechanism for heterogeneous sensors.

[0094] 2. Pure data-driven black box models lack physical constraints and have high retraining costs.

[0095] While existing end-to-end compensation models possess strong nonlinear fitting capabilities, they are essentially pure data-driven black boxes lacking physical dynamic boundary constraints. In strong transient scenarios, these models are prone to overfitting specific training distributions, leading to numerical artifacts when facing strong impact distortions and outputting frequency response distortion waveforms that violate the physical mechanisms of the sensors. Furthermore, due to the lack of prior perception of external operating conditions and adaptive adjustment mechanisms, when field sensors are replaced or the environment changes, the model must rely on massive amounts of pairwise data for extremely time-consuming iterative retraining. The high computational cost severely restricts the rapid migration and lightweight deployment of compensation systems in variable engineering environments.

[0096] 3. External engineering knowledge is difficult to effectively utilize within existing compensation frameworks.

[0097] In actual testing, the sensor's instruction manual, static calibration sheet, experimental procedures, and link configuration documents clearly record key information that determines the system's dynamic characteristics, such as nominal bandwidth, time constant, sensitivity, and operating temperature. However, existing dynamic compensation calculation frameworks can typically only process purely numerical time-series signals, and this textual engineering prior knowledge is often only used as a manual reference. Currently, there is no mechanism to effectively parse this textual information and transform it into modulation parameters that can be called upon within the compensation network, resulting in a large amount of valuable prior information not being used in actual dynamic compensation calculations.

[0098] 4. The ability to adapt and migrate to new equipment and changing operating conditions is relatively weak.

[0099] In practical engineering applications, compensation models need to be able to cope with unknown sensor models or entirely new testing environments. Existing methods, when faced with equipment replacement or significant changes in operating conditions, often require retraining or parameter calibration by collecting large amounts of experimental data due to the lack of an effective mechanism to dynamically adjust model parameters by incorporating external knowledge. This results in high migration costs for the model when facing new scenarios, making it difficult to achieve rapid adaptation and deployment across sensors and operating conditions under small sample conditions.

[0100] To address the aforementioned problems, this invention proposes a dynamic compensation method for heterogeneous sensors based on prior knowledge and conditional parameters generated from a large language model. This method incorporates the semantic parsing capabilities of a large language model for unstructured engineering knowledge into a temporal computation framework, extracting and generating the conditional parameters required for the dynamic compensation network. Simultaneously, it constructs a collaborative framework combining physical coarse compensation and residual network fine-tuning. This invention aims to achieve a substantial mapping from textual prior information to the network's internal control parameters through this mechanism, effectively improving the model's adaptive and unified compensation capabilities for multi-source heterogeneous sensors under varying operating conditions while ensuring the physical rationality of the compensation model.

[0101] The innovative aspects of this invention are as follows:

[0102] 1. A text knowledge prior extraction mechanism for dynamic compensation of heterogeneous sensors is proposed.

[0103] To address the limitations of existing compensation models that primarily rely on single numerical signals, this invention introduces a large language model as a text parsing module. This module performs semantic parsing and feature extraction on unstructured or semi-structured engineering texts such as equipment manuals, static calibration sheets, and experimental procedures, constructing computationally readable prior feature vectors. This mechanism transforms engineering text information into machine-readable numerical prior representations, providing contextual constraints that reflect the physical properties of sensors and the operating conditions for a unified compensation framework.

[0104] 2. A mechanism for generating conditional parameters of compensation networks based on prior knowledge is proposed.

[0105] This invention constructs a condition parameter generation module to map the extracted prior feature vectors and structured metadata into condition parameters (i.e., channel-level scaling and offset coefficients for performing feature affine transformations) required for the operation of the compensation network. This mechanism enables the compensation network to dynamically adjust its internal feature extraction weights and activation gating based on external prior knowledge, the physical characteristics of the current sensor, and the testing conditions. This effectively improves the model's adaptability and transfer efficiency when facing different sensor models and changing operating conditions.

[0106] 3. A unified compensation framework combining physical coarse compensation and residual correction is proposed.

[0107] This invention employs a collaborative fusion architecture in the compensation execution process: First, a physical inverse model is used to perform preliminary recovery of the measured signal based on dynamic mechanisms to ensure the causality, stability, and physical rationality of the compensation results. Then, a time-series residual network dynamically controlled by the aforementioned "condition parameters" is used to correct nonlinear dynamic errors and local transient distortions not fully covered by the physical model. This framework effectively combines the clear system boundaries of the physical model with the nonlinear fitting advantages of deep neural networks, achieving a complementary advantage of strict physical constraints and high-precision repair of complex dynamic errors.

[0108] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0109] Based on the same inventive concept, this application also provides a heterogeneous sensor dynamic compensation device based on large model knowledge priors and conditional parameters. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the heterogeneous sensor dynamic compensation device based on large model knowledge priors and conditional parameters provided below can be found in the limitations of the heterogeneous sensor dynamic compensation method based on large model knowledge priors and conditional parameters above, and will not be repeated here.

[0110] In one embodiment, a heterogeneous sensor dynamic compensation device based on large model knowledge priors and condition parameters is provided, comprising: an offline prior configuration module for acquiring multi-source data of the target sensor, including meta-information, knowledge prior vectors, and sample difference descriptors; wherein, the knowledge prior vectors are extracted from the engineering text of the target sensor, and the sample difference descriptors are used to characterize the temporal error of the target sensor;

[0111] The conditional parameter generation module is used to map multi-source data into conditional parameters for feature affine modulation.

[0112] The online execution module is used to acquire the dynamic response signal sequence of the target sensor, perform coarse compensation on the dynamic response signal sequence based on the physical model to obtain the physical coarse compensation sequence, fuse the dynamic response signal sequence and the physical coarse compensation sequence to construct a joint input feature matrix, input the joint input feature matrix into a deep neural network with conditional parameters to obtain the nonlinear residual correction signal, and obtain the dynamic compensation output signal of the target sensor based on the physical coarse compensation sequence and the nonlinear residual correction signal.

[0113] The modules in the heterogeneous sensor dynamic compensation device based on large model knowledge priors and conditional parameters can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0114] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in all of the above method embodiments.

[0115] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in all of the above method embodiments.

[0116] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in all of the above method embodiments.

[0117] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0118] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0119] 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.

[0120] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A dynamic compensation method for heterogeneous sensors based on large model knowledge priors and condition parameter generation, characterized in that, The method includes: Acquire multi-source data from the target sensor, including metadata, prior knowledge vectors, and sample difference descriptors; wherein the prior knowledge vectors are extracted from the engineering text of the target sensor, and the sample difference descriptors are used to characterize the temporal error of the target sensor. The multi-source data is mapped to conditional parameters of feature affine modulation; The dynamic response signal sequence of the target sensor is obtained, and coarse compensation is performed on the dynamic response signal sequence based on the physical model to obtain the physical coarse compensation sequence. The dynamic response signal sequence and the physical coarse compensation sequence are fused to construct a joint input feature matrix. The joint input feature matrix is ​​then input into the deep neural network of the conditional parameters to obtain the nonlinear residual correction signal. The dynamic compensation output signal of the target sensor is obtained based on the physical coarse compensation sequence and the nonlinear residual correction signal.

2. The method according to claim 1, characterized in that, Obtaining the prior knowledge vector includes: A large language model is used to map the unstructured engineering text into the knowledge prior vector.

3. The method according to claim 1, characterized in that, The conditional parameters for mapping the multi-source data to feature affine modulation include: The meta-information, the knowledge prior vector, and the sample difference descriptor are concatenated into a composite prior representation vector, which is then input into a multilayer perceptron to obtain the conditional parameters.

4. The method according to claim 1, characterized in that, The step of performing coarse compensation on the dynamic response signal sequence based on the physical model to obtain the physical coarse compensation sequence includes: Based on the meta-information and the prior knowledge vector, the physical inverse model of the target sensor is established; Based on the physical inverse model, a regularization penalty factor is introduced in the frequency domain to construct a regularized inverse operator; The inverse regularization operator is converted into a time-domain FIR filter system, and the dynamic response signal is coarsely compensated by the FIR filter system to obtain the physical coarse compensation sequence.

5. The method according to claim 1, characterized in that, The deep neural network includes a plurality of stacked residual blocks, each of which includes a one-dimensional causal dilated convolutional layer; The step of inputting the joint input feature matrix into a deep neural network to obtain the nonlinear residual correction signal includes: After each residual block completes the convolution operation, the conditional parameters are loaded to perform a channel-level feature affine transformation.

6. The method according to any one of claims 1 to 5, characterized in that, The method further includes: For changes in the target sensor type or operating conditions, a few-sample transfer learning is performed; wherein, the few-sample transfer learning includes cutting off the updates of the feature extraction convolution kernels shared across tasks in the physical model and the deep neural network, and opening up the updates of the model weights used to generate the condition parameters.

7. A dynamic compensation device for heterogeneous sensors based on large model knowledge priors and condition parameter generation, characterized in that, The device includes: An offline prior configuration module is used to acquire multi-source data from the target sensor, including metadata, knowledge prior vectors, and sample difference descriptors; wherein, the knowledge prior vectors are extracted from the engineering text of the target sensor, and the sample difference descriptors are used to characterize the temporal error of the target sensor; A conditional parameter generation module is used to map the multi-source data into conditional parameters for feature affine modulation. An online execution module is used to acquire the dynamic response signal sequence of the target sensor, perform coarse compensation on the dynamic response signal sequence based on a physical model to acquire a physical coarse compensation sequence; fuse the dynamic response signal sequence and the physical coarse compensation sequence to construct a joint input feature matrix, input the joint input feature matrix into a deep neural network of the conditional parameters to acquire a nonlinear residual correction signal; and acquire the dynamic compensation output signal of the target sensor based on the physical coarse compensation sequence and the nonlinear residual correction signal.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images