Reference construction method and device for multi-source transient signal data set

Abstract

The invention discloses a reference construction method and device for a multi-source transient signal data set, and relates to the field of sensors. Through unified scene design and multi-source sensor configuration, the system depicts dynamic behaviors of transient signals under different working conditions; a multi-dimensional dynamic index system covering core elements such as sampling precision, transient response time, dynamic range, frequency domain characteristic recovery and time sequence consistency is constructed, and quantitative evaluation of multi-source transient signal data quality is achieved; a Pareto multi-objective optimization and reinforcement learning mechanism is introduced, multi-objective equilibrium evaluation and dynamic updating of a data set benchmark are achieved, and migration and use among different tasks and different working conditions are supported; through standardized index definition, a calculation mode and an evaluation process, a set of repeatable, extensible and engineering-landing multi-source transient signal data set reference system is formed, and a unified data basis is provided for development and evaluation of a subsequent dynamic compensation algorithm and an intelligent diagnosis model.

Description

Technical Field

[0001] This application relates to the field of sensor technology, and in particular to a method and apparatus for constructing a benchmark for a multi-source transient signal dataset. Background Technology

[0002] With the continuous advancement of a new round of technological revolution and industrial transformation, high-tech fields such as aerospace, intelligent vehicles, and the Internet of Things are booming, leading to an explosive growth in demand for high-precision testing and monitoring. In these fields, the accurate capture of transient signals has become a key core for detecting rapid changes in dynamic environments. This technology is widely used in many important aspects such as high-frequency vibration testing, dynamic fault diagnosis, and performance evaluation of complex systems. In typical aerospace missions such as launch vehicle launches, engine ground hot-fire tests, and rocket flight attitude maneuvers, structural vibrations, impact loads, cabin pressure fluctuations, and multi-physics coupling effects change rapidly within milliseconds, placing extremely high demands on the dynamic response performance of sensors and the consistency of testing systems. The accuracy of transient test data directly affects flight status interpretation, fault location, and the reliability of subsequent model improvement designs. The launch and explosion processes of weapons release signals with a large dynamic range in an extremely short time, placing stringent requirements on the sensitivity, response speed, and anti-interference capabilities of sensors, directly affecting the safety and effective functioning of weapon systems.

[0003] In current engineering practice, transient testing in the aerospace field relies heavily on a large number of distributed heterogeneous sensors for pressure, acceleration, strain, and temperature. These sensors differ significantly in dynamic range, frequency response, noise characteristics, and environmental adaptability, leading to significant inconsistencies in multi-source transient signals under the same operating conditions. Furthermore, the complex environment of launch sites—high temperature, high humidity, and high sound pressure levels—coupled with the complex conditions of high overload, high vibration, and high coupling during flight, results in transient events characterized by "short-term bursts, strong nonlinearity, and strong coupling," making test data highly susceptible to distortion, drift, and loss in the time-frequency domain. While a few open-source shock and vibration datasets exist in the aerospace field, they are mostly historical records from specific test missions, with limited access and incomplete condition annotations. Moreover, existing datasets tend to focus on single sensors or single scenarios, failing to meet the needs for unified compensation for heterogeneous sensors and cross-scenario algorithm evaluation.

[0004] To support flight tests and ground verification missions, several transient signal datasets tailored to specific missions have been gradually accumulated in engineering practice, such as long-duration test datasets for a certain type of launch vehicle engine and structural vibration test datasets. In addition, there are publicly available aerospace shock signal datasets, represented by NASA (National Aeronautics and Space Administration) shock test data, which are widely used for structural dynamics analysis and fault diagnosis algorithm verification. However, these datasets are usually collected under specific models and conditions, with limited scenario coverage, making them difficult to directly transfer to other missions or new configurations. Most datasets focus on single-sensor data, lacking a data system that integrates multi-source and multi-physical quantity acquisition. Dataset construction emphasizes data acquisition itself, lacking systematic and quantifiable benchmark definitions for indicators such as sampling accuracy, transient response, dynamic range, frequency domain recovery, and cross-sensor temporal consistency. Given the scarcity of open-source data, researchers often rely on building their own datasets and using methods such as generative adversarial networks and transfer learning for data augmentation to alleviate the problems of insufficient samples and imbalanced distribution. However, such methods generally reuse existing data, making it difficult to fundamentally solve problems such as discrepancies between synthetic data and real-world conditions, and the lack of unified evaluation standards for data quality assessment.

[0005] In summary, existing methods for constructing benchmark datasets have limited application scope, especially in complex aerospace applications. These methods cannot be directly used as benchmarks for constructing multi-source sensor datasets, making it difficult to provide an effective basis for constructing high-quality datasets required for transient testing. Therefore, there is an urgent need to develop a multi-dimensional benchmark dataset construction method that can accommodate multi-source heterogeneous sensors, complex transient conditions, and engineering reliability requirements, in order to support the reliable training and objective evaluation of subsequent dynamic compensation, fault diagnosis, and health assessment algorithms. Summary of the Invention

[0006] Therefore, it is necessary to provide a benchmark construction method and apparatus for multi-source transient signal datasets that can provide higher quality benchmark datasets, addressing the aforementioned technical problems.

[0007] Firstly, this application provides a benchmark construction method for a multi-source transient signal dataset. The method includes:

[0008] Acquire multi-source transient signals under transient conditions in a single scenario;

[0009] Construct a multi-dimensional dynamic index system, map multi-source transient signals to the quality space under the multi-dimensional dynamic index system, and obtain a heterogeneous index set;

[0010] A Pareto multi-objective optimization model is constructed, and a candidate dataset for the multi-objective optimization model is obtained based on a heterogeneous index set. A non-dominated solution set is obtained through a non-dominated sorting genetic algorithm. The non-dominated solution set is comprehensively scored, and a reference benchmark for multi-source transient signals is selected from the non-dominated solution set based on the scoring results.

[0011] The reference benchmark is dynamically updated through reinforcement learning.

[0012] In one embodiment, the multidimensional dynamic index system includes indicators characterizing sampling accuracy, transient response time, dynamic range, frequency domain characteristic recovery, and timing consistency.

[0013] In one embodiment, obtaining a candidate dataset for a multi-objective optimization problem based on a heterogeneous index set includes:

[0014] Based on the polarity of indicators in the multidimensional dynamic indicator system, the heterogeneous indicator set is dimensionless to obtain candidate datasets; among them, the polarity of indicators includes benefit-type indicators and cost-type indicators.

[0015] In one embodiment, a comprehensive score for the non-dominated solution set includes:

[0016] Using geometric mean as the decision function, a comprehensive scoring function is constructed, and the scoring results of the non-dominated solution set are obtained using the comprehensive scoring function.

[0017] In one embodiment, dynamically updating a reference benchmark through reinforcement learning includes: constructing a state space and an action space;

[0018] The state space consists of a heterogeneous set of indicators and task preference weights; the action space includes actions that intervene in the multi-source transient signal acquisition process and the candidate dataset acquisition process.

[0019] In one embodiment, dynamically updating the reference benchmark through reinforcement learning further includes: constructing a reward function;

[0020] The reward function is constructed based on the comprehensive scoring function, the dynamic scoring baseline, and the cost function; the cost function is constructed based on the resource consumption caused by the execution action.

[0021] Secondly, this application also provides a benchmark construction apparatus for a multi-source transient signal dataset. The apparatus includes:

[0022] The offline training module is used to acquire multi-source transient signals under single-scenario transient conditions; construct a multi-dimensional dynamic index system, map the multi-source transient signals to the quality space under the multi-dimensional dynamic index system, and obtain a heterogeneous index set; construct a Pareto multi-objective optimization model, and obtain a candidate dataset for the multi-objective optimization model based on the heterogeneous index set, and obtain a non-dominated solution set through a non-dominated sorting genetic algorithm; perform a comprehensive score on the non-dominated solution set, and select a reference benchmark for multi-source transient signals from the non-dominated solution set based on the score results;

[0023] The online fine-tuning module is used to dynamically update the reference benchmark through reinforcement learning.

[0024] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps in the aforementioned benchmark construction method for multi-source transient signal datasets.

[0025] 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 aforementioned benchmark construction method for multi-source transient signal datasets.

[0026] 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 aforementioned benchmark construction method for multi-source transient signal datasets.

[0027] The aforementioned benchmark construction method and apparatus for multi-source transient signal datasets, through unified scenario design and multi-source sensor configuration, systematically characterizes the dynamic behavior of transient signals under different operating conditions; constructs a multi-dimensional dynamic index system covering core elements such as sampling accuracy, transient response time, dynamic range, frequency domain characteristic recovery, and temporal consistency, enabling quantitative evaluation of multi-source transient signal data quality; introduces Pareto multi-objective optimization and reinforcement learning mechanisms to achieve multi-objective balanced evaluation and dynamic updating of the dataset benchmark, supporting migration and use across different tasks and operating conditions; and through standardized index definitions, calculation methods, and evaluation processes, forms a repeatable, scalable, and engineering-applicable multi-source transient signal dataset benchmark system, providing a unified data foundation for the development and evaluation of subsequent dynamic compensation algorithms and intelligent diagnostic models. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating a baseline construction method for a multi-source transient signal dataset in one embodiment;

[0029] Figure 2 This is a block diagram illustrating the construction principle of a multidimensional dynamic indicator system in one embodiment;

[0030] Figure 3 This is a schematic diagram of the process for obtaining a standard reference set in one embodiment;

[0031] Figure 4 This is a schematic diagram of the principle of dynamic updating of a standard reference set based on reinforcement learning in one embodiment. Detailed Implementation

[0032] 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.

[0033] Sensors are subject to slight deviations due to factors such as manufacturing, aging, and temperature changes. By periodically or initially setting a reference, sensors can be calibrated to correct these inherent errors and ensure the accuracy of sensor measurements.

[0034] However, existing methods for acquiring sensor reference data often suffer from the following problems:

[0035] 1. Existing benchmarks are mostly based on static data and are designed for single or a small number of sensor types, making it difficult to characterize the consensus response of multiple sensor sources to the same transient event under aerospace transient conditions.

[0036] 2. Existing datasets are mostly based on specific historical missions or simple laboratory environments, and do not adequately cover typical aerospace operating conditions such as high temperature and humidity, strong sound and vibration, and violent pressure fluctuations at launch sites, resulting in insufficient transferability of datasets when used across operating conditions.

[0037] 3. Existing evaluation indicators are mostly focused on static error and noise level, lacking quantitative consideration of key dynamic indicators such as transient response time, dynamic range, frequency domain characteristic recovery, and multi-sensor timing consistency, and thus failing to characterize data quality from multiple perspectives.

[0038] 4. Current dataset evaluations often use static weighting, which makes it difficult to take into account multiple performance indicators such as sampling accuracy, response speed, dynamic range, frequency domain consistency, and time series consistency. Moreover, once the benchmark is set, it is difficult to update it automatically with the needs of new tasks.

[0039] To address the aforementioned problems, embodiments of this application provide a benchmark construction method for multi-source transient signal datasets, such as... Figure 1 As shown, it includes the following steps:

[0040] Step 101: Obtain multi-source transient signals under a single-scenario transient condition.

[0041] Targeting typical transient operating conditions such as aerospace engine testing, structural impact testing, and launch process environmental monitoring, this project constructs a representative set of operating conditions by planning multiple experimental scenarios covering extreme environments (high temperature, low temperature, high pressure) and typical interferences (noise interference, data loss). The scenario design follows the principle of "single basic environment + single typical interference" for controlled variables: the basic environment includes high temperature, low temperature, and high pressure environments, simulating typical aerospace operating conditions such as high temperature at the launch site, high-altitude low temperature, and pressure fluctuations in the launch compartment; the interference scenarios include bandwidth-limited noise interference and data loss interference, simulating measurement noise in the complex environment of the launch site and packet loss phenomena that may occur in the measurement link. By logically combining the basic environment and typical interference, several representative complex operating condition scenarios are constructed, covering the main environmental stresses and interference types in typical aerospace missions while ensuring a controllable number of scenarios.

[0042] Specifically, in terms of scenario design, to address the diverse environmental requirements of multi-sensor testing under transient shocks, an experimental scenario system covering multiple environments and interferences is constructed. This system combines collaborative data acquisition from multiple types of sensors, clearly defining the required physical parameter ranges and control logic for each scenario to ensure data representativeness and reproducibility. It covers the collaborative and consistent responses of multiple sensors in complex environments. Scenario classification is based on two dimensions: extreme environments and interference scenarios, aiming to simulate the complex conditions sensors may face in the real world. Extreme environments primarily verify the sensor's basic measurement capabilities and stability by simulating sensor performance under extreme physical conditions. Interference scenarios focus on examining common interference factors encountered by sensors during data transmission and signal acquisition, evaluating their robustness and consistency.

[0043] In each scenario, the basic principle of "multi-sensor collaborative acquisition in a single scenario" is followed. Using the same physical event as a reference, multiple sensors, including those for pressure, acceleration, temperature, and strain, are simultaneously excited and acquired. Sensor selection revolves around the key characteristics of transient signals. Taking the aerospace field as an example, in high-pressure scenarios, piezoelectric pressure sensors with a range of 0-20 MPa and a frequency response of 0-1 kHz are used to capture dynamic changes in chamber pressure fluctuations and impact pressure. In impact and vibration scenarios, piezoelectric accelerometers with a frequency upper limit of 10 kHz and a sensitivity of 100 mV / g are used to capture high-frequency vibration information. Simultaneously, foil strain sensors and thermocouple temperature sensors are configured to jointly observe the same transient event from both structural deformation and thermal environment perspectives. Through this multi-source configuration, this invention constructs "multi-view observation of the same transient event under different physical quantities" in each scenario, providing fundamental data for subsequent analysis of multi-source consistency and dynamic characteristics.

[0044] Based on the scenario testing objectives, sensors covering key signal dimensions were selected. Key parameters of the selected sensors, such as range, operating temperature, and frequency response, were carefully designed to ensure they meet environmental adaptability and signal response range matching requirements based on transient signal characteristics, covering the signal dimensions needed for dynamic compensation tasks. To ensure strict alignment of multi-sensor data in the time dimension, this invention employs a unified synchronous acquisition controller to provide a unified sampling clock and trigger signal for all channels, controlling cross-channel time synchronization errors to the microsecond level. At least 100 valid samples are collected for each scenario, with each sample recording length of 10 seconds, including a stable baseline segment of approximately 2 seconds before the event trigger and a complete response segment of approximately 8 seconds after the trigger. Baseline data can be used for sensor zero-point calibration and noise level estimation, while response data is used for transient characteristics and multi-source consistency evaluation. All data is stored in binary format and retains high-precision timestamp information, thus providing a traceable data foundation for subsequent index calculations, repeated experiments, and third-party verification. Through the above-mentioned scenarios and data acquisition design, this invention organically combines scenarios, sensors, and data acquisition strategies while maintaining the aerospace application background, providing a scientific, reasonable, and reproducible experimental basis for the subsequent construction and verification of benchmark indicator systems.

[0045] Step 102: Construct a multi-dimensional dynamic index system, map multi-source transient signals to the quality space under the multi-dimensional dynamic index system, and obtain a heterogeneous index set.

[0046] In multi-source transient signal datasets, the signal characteristics of different sensors and different scenarios vary significantly. To establish a unified and quantifiable benchmark for data quality evaluation, this invention proposes a five-dimensional benchmark index system consisting of sampling accuracy, transient response time, dynamic range, frequency domain characteristic recovery, and temporal consistency. This system covers both single-sensor measurement accuracy and dynamic consistency among multiple sensors, comprehensively characterizing dataset quality from three dimensions: time domain, frequency domain, and multi-source collaboration.

[0047] This invention represents dataset quality as:

[0048]

[0049] in Characterizes sampling accuracy. Characterizing transient response time, Characterizes dynamic range, Characterizes the ability to recover frequency domain characteristics. Characterizing temporal consistency. This invention maps heterogeneous data from different sources to the same quality space through unified quantitative definitions and threshold design.

[0050] Specifically, such as Figure 2The diagram shown illustrates the construction of a multidimensional dynamic index system. Sampling accuracy is a fundamental indicator for measuring a sensor's ability to reproduce the static or quasi-static components of a physical quantity. This invention uses two sub-indicators—average relative error and resolution—for joint characterization. Average relative error... Used to assess how much a measured value deviates from the true value. Assumptions The measurement value at the k-th sampling point, Let N be the reference true value corresponding to the same moment. This true value is obtained through parallel acquisition by a high-precision standard sensor or from a high-confidence calibration condition. N is the number of effective measurement points. The average error is defined as follows:

[0051]

[0052] resolution Used to evaluate the system's ability to detect subtle signal changes. Let... Let n be the full-scale range of the sensor, and n be the number of bits in the A / D conversion of the data acquisition system. Then, we define:

[0053]

[0054] Setting a benchmark judgment logic: To ensure the comparability of multi-source sensors in terms of accuracy, this invention sets a unified qualified threshold, requiring an average relative error. and resolution It should not exceed 1 / 2 of the minimum effective change in the measured signal.

[0055] In aerospace shock and explosion tests, signals have extremely fast rise times. Transient response time is a key consideration of the sensor's ability to dynamically track and capture transient abrupt changes. For the abrupt change moment, a first-order differential abrupt change detection algorithm is employed. Specifically, for the sensor output signal... Find the derivative (or difference). Calculate the standard deviation of the baseline noise. When the rate of change of the signal first exceeds the set noise threshold T (take... When this happens, mark that moment as the start of the signal mutation. :

[0056]

[0057] Assuming the steady-state value (or peak value) of the signal after the abrupt change is Define response time. From the moment of mutation onset From the start until the output signal first reaches 90% of its steady-state value. The time interval between:

[0058]

[0059] Based on the physical characteristics of typical transient events in aerospace, this invention imposes mandatory constraints. It must be less than 1 / 10 of the event characteristic pulse width. For example, for an impact event with a pulse width of 0.5 ms, the required response time of the accelerometer is... This ensures that the data accurately records the steepness of the area before the shock wave, avoiding waveform distortion caused by sensor response lag.

[0060] Dynamic range is used to evaluate a sensor's ability to withstand strong peak impacts while clearly distinguishing weak signal details under the same operating conditions. It is expressed in decibels (dB) as the ratio of the minimum to the maximum signal that the sensor can reliably measure. Let... This represents the maximum effective amplitude within the full-scale range of the sensor's linear operating region. To achieve the preset signal-to-noise ratio threshold (in this embodiment, we take...) The smallest resolvable signal amplitude is then:

[0061]

[0062] To address the high dynamic characteristics during aerospace launch and separation processes, this invention establishes a classification criterion: requiring piezoelectric accelerometers to... Pressure sensor This ensures that strong signal clipping is avoided while weak signal details are preserved within the same dataset.

[0063] Since transient signals often contain wideband information, frequency domain characteristic recovery metrics are used to quantify the consistency of spectral feature reconstruction by multi-source sensors within their main operating frequency bands. Let... Let f be the amplitude-frequency response function of the i-th sensor to be evaluated at frequency f. This refers to the amplitude-frequency response of a standard reference sensor. Within the operating frequency band of interest. Internally, define the normalized frequency domain consistency error. :

[0064]

[0065] In engineering implementation, the integral is approximated by summing discrete frequency points. An error threshold is set. This constraint ensures that, during multi-source data fusion analysis, different sensors provide consistent spectral characterization of the same transient event within their main operating frequency band.

[0066] To address the needs of multi-source data fusion and causal analysis, this indicator imposes dual constraints from two dimensions: "absolute time synchronization" and "relative waveform smoothness." Maximum timestamp deviation. This is used to measure the time alignment accuracy between different sensor channels. Assume the timestamps of the i-th sensor and the j-th sensor at the k-th sampling point are respectively... and The maximum timestamp deviation between the two sensors is defined as:

[0067]

[0068] This invention requires (Preferred) This is to meet the correlation analysis requirements of millisecond-level transient events. And regarding signal timing smoothness... This invention introduces a simple first-order differential smoothness index to eliminate non-physical jumps caused by transmission packet loss, timing jitter, or anomalies. Mathematically defined as:

[0069]

[0070] Where M is the number of sample points. Set constraints for steady-state values. This ensures strict alignment and sufficient smoothness of multi-source data on the time axis, providing a reliable prerequisite for multi-source fusion and dynamic compensation algorithms.

[0071] By constructing the above five indicators, this invention maps multi-source heterogeneous transient data, which were originally difficult to compare directly, into a computable, interpretable, and comparable five-dimensional quality space Q. This not only provides a standard for judging whether a single dataset is qualified, but also provides a rigorous quantitative input (State) and reward function (Reward) foundation for Pareto optimization-based dataset selection and reinforcement learning-based benchmark dynamic updates in subsequent embodiments.

[0072] Step 103: Construct a Pareto multi-objective optimization model and obtain candidate datasets for the multi-objective optimization model based on the heterogeneous index set. Obtain non-dominated solution set through non-dominated sorting genetic algorithm. Perform comprehensive scoring on the non-dominated solution set and select a reference benchmark for multi-source transient signals from the non-dominated solution set based on the scoring results.

[0073] In aerospace multi-source transient signal testing, performance indicators of different dimensions often exhibit physical incompatibilities. For example, highly sensitive sensors are often accompanied by narrow bandwidths or slow transient responses. To find the optimal dataset construction scheme among these conflicting objectives, this invention proposes an evaluation method based on Pareto front optimization and geometric mean weighted decision-making.

[0074] like Figure 3As shown, firstly, the five-dimensional indicators constructed in step 102 have different physical dimensions (such as time in milliseconds, dynamic range in dB, and error in %) and inconsistent optimization directions. Therefore, this invention first establishes a unified mapping space. Assume the original indicator set is a vector. These correspond to sampling accuracy, transient response time, dynamic range, frequency domain characteristic recovery, and timing consistency, respectively. Based on the nature of the indicators, they are divided into "benefit-oriented indicators" (the larger the better) and "cost-oriented indicators" (the smaller the better).

[0075] set up and These are the minimum and maximum values ​​of the indicator in the current candidate dataset, respectively. For benefit-type indicators (e.g. The dynamic range is mapped to the interval [0,1] using a standard Min-Max linear transformation:

[0076]

[0077] For cost-related indicators (e.g.) relative error Response time Frequency domain error, The timing deviation is addressed using an inverse Min-Max transformation, which not only eliminates the dimensions but also unifies them into a "larger is better" principle through the inverse transformation.

[0078]

[0079] After the above processing, each candidate dataset scheme k is transformed into a dimensionless five-dimensional mass vector. ,in Furthermore, a larger value indicates better quality in that dimension. This vector serves as a direct input variable for subsequent multi-objective optimization.

[0080] This invention seeks the "optimal balance" of multidimensional indicators, and models the dataset benchmark selection as a multi-objective maximization problem, thus constructing an objective function vector:

[0081]

[0082] In this model, Pareto Dominance is defined as follows: for any two datasets, schemes A and B, if scheme A performs no worse than scheme B in all five dimensions (i.e., ... ), and is strictly superior to B in at least one dimension (i.e. If the candidate set..., then scheme A dominates scheme B. The scheme in Not If it is dominated by any other scheme, then it is called The set of all Pareto optimal solutions is called the Pareto Front. This front represents the physical limit of achievable multi-source transient signal quality under current sensor hardware limitations and environmental interference.

[0083] Given the enormous search space for dataset construction schemes, this invention employs a non-dominated sorting genetic algorithm (NSGA-II) with an elitist strategy for efficient solution. The specific process is as follows:

[0084] Step 1031, Fast Non-dominated Sorting. The algorithm stratifies the population based on dominance relationships. First, it identifies all non-dominated individuals to form the first front. (Optimal level); Remove from the population Then, the non-dominated solutions among the remaining individuals are searched to form the second frontier. This process, in turn, ensures that the selection process strictly adheres to the principle of prioritizing "multi-dimensional comprehensive quality".

[0085] Step 1032, Crowding Distance Computation. To avoid the solution set getting trapped in local optima, such as extreme solutions with only high dynamic range but extremely slow response, this invention introduces the crowding distance calculation. For individuals within the same frontier, calculate their Euclidean distance to neighboring individuals in the target space:

[0086]

[0087] Algorithm priority retention Larger individuals force the solution set to be evenly distributed on the Pareto front, ensuring that the selected benchmark covers operating conditions with different preferences.

[0088] Step 1033, Evolution and Iteration. Offspring are generated through selection, crossover, and mutation operations, and the parent and offspring are merged. Using the aforementioned sorting and crowding selection mechanism, the optimal N individuals are selected to enter the next generation, until the preset number of iterations is reached or the frontier converges.

[0089] While the Pareto front provides a set of nondominated solutions, engineering applications ultimately require the determination of specific baseline values. This invention employs the geometric mean as the decision function to construct a comprehensive scoring function. :

[0090]

[0091] in Weighting coefficients (default) The effectiveness of geometric mean lies in its inherent balance constraint mechanism. If a dataset scheme has a serious weakness in any dimension (e.g., ...), it will mitigate the impact of geometric mean. Even if other indicators are excellent, the overall product will tend to be 0. Furthermore, compared to the arithmetic mean, the geometric mean has a stronger penalizing effect on highly discrete data sets. The benchmark dataset selected by this mechanism meets the required standards in five aspects, including sampling accuracy, response speed, and dynamic range, thus ensuring its authority and robustness as a "benchmark." Ultimately, according to... The data were sorted, and the Top-K scheme was selected as the reference benchmark set for the multi-source transient signal dataset.

[0092] Step 104: Dynamically update the reference benchmark through reinforcement learning.

[0093] Traditional dataset benchmark construction typically employs fixed indicator weights from the initial design and maintains them unchanged over the long term. This static mechanism is ill-suited to the highly time-varying mission requirements of aerospace testing. In different testing phases, some tasks focus on capturing full-band signals, emphasizing transient response time and frequency domain characteristic recovery; while others prioritize preserving key data features, focusing on sampling accuracy and temporal consistency. To resolve the contradiction between static benchmarks and dynamic mission requirements, this invention proposes a Pareto-guided, reinforcement learning-driven adaptive benchmark update framework. This framework does not rely on manually preset fixed rules but models the dataset benchmark maintenance process as a Markov Decision Process (MDP). Through interaction between an intelligent agent and the testing environment, it perceives the current data state and mission preferences, dynamically adjusting acquisition and filtering strategies accordingly. This enables the benchmark system to possess adaptive adjustment and continuous optimization capabilities.

[0094] like Figure 4 As shown, in this decision-making process model, the present invention first constructs a state space that includes both intrinsic data quality and extrinsic task orientation. Specifically, the state vector is defined as a combination of the normalized quality vector of the current data segment or dataset scheme and task preference information, i.e. .in, The normalized index value of the selected scheme in the i-th dimension at the current time t (as described in Section 3.3) reflects the quality level of the current data in real time. Task preference weights are dynamically issued from upper-level task instructions, reflecting the current task's focus on different indicators. For example, in a high-reliability health monitoring task, the weights for sampling accuracy and timing consistency can be increased. Simultaneously, this invention defines an action space capable of directly intervening in the data generation and processing flow, containing a series of discrete adjustment strategies: at the acquisition level, these include increasing or decreasing the sampling frequency, switching between high-sensitivity and low-power modes, etc., to physically alter the quality of the original signal; at the processing level, these include tightening or relaxing the screening thresholds for different indicators, selecting which segments to retain, marking them as abnormal, or triggering data repair. By selecting different strategies within the action space, the agent can indirectly change the distribution of future data samples in the five-dimensional quality space, thereby guiding the benchmark to gradually evolve in a direction more aligned with the current task requirements.

[0095] To guide the intelligent agent in finding the optimal balance between pursuing high-quality data and controlling system resource consumption, the design of the reward function constitutes the core innovation of this invention. Based on the Pareto frontier and geometric mean combined score calculated in step 103, this invention defines the immediate reward at each step as:

[0096]

[0097] in, The data fragments or schemes obtained after adopting the current strategy. The overall score. It should be noted that, in order to achieve task adaptability, the exponential weights in the formula are adjusted when calculating this score. The task preference weights in the state space will be used directly. Dynamic values ​​are assigned to represent the quality of the data for the current specific task; This is a dynamic scoring baseline. This parameter is not a fixed value, but a sliding baseline calculated based on the mean or a specific quantile of the scoring results within a historical time window, representing the recent "average level". In order to take action The cost function is used to quantify the resource consumption caused by performing the action, such as the storage pressure or bandwidth usage caused by high sampling frequency; An adjustment factor used to balance performance and cost.

[0098] Through this differentiated reward design, when the agent's chosen actions result in data quality significantly higher than the current dynamic scoring baseline, When this happens, a significant positive reward will be obtained; conversely, if an agent single-mindedly pursues improvement in a single metric, causing other metrics to deteriorate severely, resulting in a lower weighted score S, or leading to excessive resource consumption, a significant positive reward will be obtained. If the score is too high, the overall score will decrease or the cost will increase, and the reward will decrease accordingly. This mechanism forces reinforcement learning agents to actually learn how to find the optimal data construction strategy under multiple objective trade-offs and engineering resource constraints, with the Pareto front as the quality boundary.

[0099] In terms of specific engineering implementation and training process, this invention adopts a two-stage strategy of "offline pre-training + online fine-tuning". First, a simulation environment is constructed using historically accumulated data, and offline training is performed using Deep Q-Learning (DQN) or a policy gradient-based method. During this stage, the agent continuously tries different combinations of sampling and selection strategies, observes their performance in the five-dimensional quality space, and updates the network parameters using the Bellman equation, learning a set of initial strategies that maximize the overall quality of the dataset in the long-term reward context. Subsequently, the trained agent is deployed to actual test tasks for online fine-tuning. With the continuous addition of new task scenarios and newly collected data, the system periodically recalculates the Pareto frontier using the new data and updates the reference benchmark. The computational logic is then used to continue training the agent to adapt to the new task preference weights. This mechanism solves the scientific problems in traditional benchmark design, such as the difficulty in unifying index weights and the inability of benchmarks to adapt to changes in missions. It realizes the transformation from a static, manually set benchmark system to a dynamic, data-driven adaptive benchmark system, and provides methodological support for the reliable migration and long-term evolution of aerospace multi-source transient signal datasets across different missions and platforms.

[0100] In one embodiment, the overall process of constructing a benchmark for a multi-source transient signal dataset includes:

[0101] Step 201 involves collecting sensor data from various sensors commonly used in aerospace transient testing, such as those for pressure, acceleration, temperature, and strain. A five-dimensional benchmark index system is proposed, comprising sampling accuracy, transient response time, dynamic range, frequency domain characteristic recovery, and temporal consistency. This system covers both single-sensor measurement accuracy and dynamic consistency among multiple sensors, comprehensively characterizing dataset quality from three dimensions: time domain, frequency domain, and multi-source collaboration.

[0102] Step 202 addresses the trade-offs among the five dimensions (such as the conflict between accuracy and response time, and dynamic range and noise suppression). A Pareto multi-objective optimization model is constructed, and normalization methods such as Min-Max or Z-score are used to eliminate dimensional differences among the indicators. The dataset or data segment is considered as a candidate solution to the multi-objective optimization problem. A non-dominated sorting genetic algorithm (such as NSGA-II) is used to solve the Pareto front, selecting the non-dominated solution set that demonstrates balanced performance across multiple indicators. Based on this, unbiased synthesis methods such as geometric mean are used to evaluate the Pareto front solutions, achieving objective ranking and selection of dataset quality.

[0103] Step 203: The multi-source transient signal "multi-dimensional index vector of the current data segment" is used as the state space of the reinforcement learning agent; "sampling mode switching, trigger strategy adjustment, abnormal segment removal / marking" are used as the action space, and the agent is guided to select the optimal strategy in the action space; the comprehensive score based on the Pareto front is used as the reward function. By training the RL agent, it can automatically adjust the attention of each index under different working conditions and task requirements, realize the adaptive adjustment of the benchmark weight and the dynamic update of the benchmark curve, and solve the limitation of traditional fixed benchmarks that are difficult to adapt to new tasks and complex working conditions.

[0104] Step 204: In combination with typical working conditions such as high temperature, low temperature, high pressure, noise interference, and data loss, design experimental scenarios covering different extreme environments and interference combinations, and deploy multiple sensors in a single scenario to achieve collaborative data acquisition. Strictly control time alignment error through hardware synchronization methods to provide high-quality raw data for the calculation and verification of benchmark indicators.

[0105] In summary, to address the lack of a unified dataset quality benchmark for various types of sensors under complex transient conditions, this invention constructs a complete technical route encompassing scenario design, collaborative data acquisition, multi-dimensional benchmark index construction, multi-objective optimization, and dynamic benchmark updating. Through this route, this invention not only provides a quantitative index system to describe the quality of multi-source transient signal datasets but also proposes an adaptive benchmark construction method that deeply integrates multi-objective optimization and reinforcement learning. This allows the benchmark to be automatically updated according to changes in task requirements and actual data distribution, thereby solving the scientific problem that existing static benchmarks are difficult to adapt to multiple operating conditions and tasks, and ensuring data accuracy under transient testing conditions.

[0106] Compared with existing methods for constructing and evaluating multi-source datasets, this invention has the following significant advantages and beneficial effects by constructing a combined technical solution of multi-scenario experimental design, multi-dimensional index system, Pareto multi-objective optimization, and reinforcement learning dynamic update.

[0107] First, this invention overcomes the limitation of existing benchmark methods that only cover a single scenario. Starting from the practical needs of aerospace transient testing, this invention designs a composite experimental scenario covering high temperature, high pressure, shock, and typical interference. This design can accurately capture and characterize the dynamic behavior and multi-source consistency of multi-source transient signals under extreme conditions, filling the gap in the industry for a dedicated dataset benchmark for multi-source heterogeneous transient signals. Simultaneously, it avoids the one-sidedness of traditional single-precision index evaluation, significantly improving the quality of the dataset through a multi-dimensional index collaborative evaluation system. Unlike existing technologies that only focus on static error or noise levels, this invention can comprehensively quantify the data from multiple dimensions, including time-domain transient tracking, frequency-domain feature reconstruction, and multi-source time-series alignment. This provides clear and interpretable quantitative basis for subsequent optimization of test schemes and data cleaning processes.

[0108] Secondly, this invention overcomes the problems of indicator imbalance and transfer difficulties that are easily caused by traditional weighted average evaluation methods. Addressing the physical contradictions between indicators such as high sensitivity and wide bandwidth, and high precision and fast response in transient testing, this invention introduces a Pareto multi-objective optimization model and a non-dominated ranking algorithm. This automatically searches the physical boundaries of each performance indicator and, combined with geometric mean decision-making, selects a balanced solution that performs harmoniously across all dimensions without significant weaknesses, ensuring the comprehensive robustness of the benchmark dataset. Furthermore, traditional benchmarks, once set, remain unchanged for a long period, making it difficult to adapt to the differentiated needs of different testing tasks at various stages. This invention models benchmark construction as a Markov decision process, using reinforcement learning agents to perceive task preferences and changes in data distribution, automatically adjusting the acquisition strategy and screening thresholds, significantly improving the benchmark's transferability and reusability across different task types, testing platforms, and testing stages.

[0109] Finally, a unified and standardized testing platform is provided for dynamic compensation and intelligent diagnostic algorithms, improving the reliability of engineering applications. Based on the high-quality benchmark dataset constructed in this invention, subsequent sensor dynamic compensation models, fault diagnosis models, and health assessment algorithms can be trained and validated under the same set of rigorous data standards and indicator frameworks. This method can eliminate algorithm evaluation bias caused by inconsistent data sources and varying data quality, enabling fair comparison and objective evaluation between different algorithms.

[0110] 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.

[0111] Based on the same inventive concept, this application also provides a reference construction apparatus for a multi-source transient signal dataset. The solution provided by this apparatus is similar to the solution described in the above method. Therefore, the specific limitations of one or more reference construction apparatus embodiments for multi-source transient signal datasets provided below can be found in the limitations of the reference construction method for multi-source transient signal datasets described above, and will not be repeated here.

[0112] In one embodiment, a benchmark construction device for a multi-source transient signal dataset is provided, comprising: an offline training module for acquiring multi-source transient signals under a single-scene transient condition; constructing a multi-dimensional dynamic index system, mapping the multi-source transient signals to a quality space under the multi-dimensional dynamic index system, and obtaining a heterogeneous index set; constructing a Pareto multi-objective optimization model, and obtaining a candidate dataset for the multi-objective optimization model based on the heterogeneous index set, and obtaining a non-dominated solution set through a non-dominated sorting genetic algorithm; performing a comprehensive score on the non-dominated solution set, and selecting a reference benchmark for the multi-source transient signal from the non-dominated solution set based on the score results;

[0113] The online fine-tuning module is used to dynamically update the reference benchmark through reinforcement learning.

[0114] Each module in the aforementioned reference construction device for multi-source transient signal datasets 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 memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0115] 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.

[0116] 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.

[0117] 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.

[0118] 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 the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0119] 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.

[0120] 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.

[0121] 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 method for constructing a benchmark for a multi-source transient signal dataset, characterized in that, The method includes: Acquire multi-source transient signals under transient conditions in a single scenario; A multi-dimensional dynamic index system is constructed, and the multi-source transient signals are mapped to the quality space under the multi-dimensional dynamic index system to obtain a heterogeneous index set. A Pareto multi-objective optimization model is constructed, and a candidate dataset for the multi-objective optimization model is obtained based on the heterogeneous index set. A non-dominated solution set is obtained through a non-dominated sorting genetic algorithm. The non-dominated solution set is comprehensively scored, and a reference benchmark for the multi-source transient signal is selected from the non-dominated solution set based on the scoring results. The reference benchmark is dynamically updated through reinforcement learning.

2. The method according to claim 1, characterized in that, The multidimensional dynamic index system includes indicators that characterize sampling accuracy, transient response time, dynamic range, frequency domain characteristic recovery, and timing consistency.

3. The method according to claim 1, characterized in that, The step of obtaining the candidate dataset for the multi-objective optimization problem based on the heterogeneous index set includes: Based on the polarity of the indicators in the multidimensional dynamic indicator system, the heterogeneous indicator set is dimensionless to obtain the candidate dataset; wherein, the polarity of the indicators includes benefit-type indicators and cost-type indicators.

4. The method according to claim 1, characterized in that, The comprehensive scoring of the non-dominated solution set includes: Using geometric mean as the decision function, a comprehensive scoring function is constructed, and the scoring result of the non-dominated solution set is obtained using the comprehensive scoring function.

5. The method according to claim 1, characterized in that, The step of dynamically updating the reference benchmark through reinforcement learning includes: constructing a state space and an action space; The state space consists of the heterogeneous index set and task preference weights; the action space includes actions that intervene in the multi-source transient signal acquisition process and the candidate dataset acquisition process.

6. The method according to claim 5, characterized in that, The step of dynamically updating the reference benchmark through reinforcement learning further includes: constructing a reward function; The reward function is constructed based on the comprehensive scoring function, the dynamic scoring baseline, and the cost function; the cost function is constructed based on the resource consumption caused by the execution action.

7. A reference construction apparatus for a multi-source transient signal dataset, characterized in that, The device includes: An offline training module is used to acquire multi-source transient signals under a single-scene transient condition; construct a multi-dimensional dynamic index system, map the multi-source transient signals to the quality space under the multi-dimensional dynamic index system, and obtain a heterogeneous index set; construct a Pareto multi-objective optimization model, and obtain a candidate dataset for the multi-objective optimization model based on the heterogeneous index set, and obtain a non-dominated solution set through a non-dominated sorting genetic algorithm; perform a comprehensive score on the non-dominated solution set, and select a reference benchmark for the multi-source transient signal from the non-dominated solution set based on the score results; An online fine-tuning module is used to dynamically update the reference benchmark through reinforcement learning.

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