Sensor testing method based on dynamic environment simulation and performance attenuation prediction

Abstract

The invention discloses a sensor testing method based on dynamic environment simulation and performance attenuation prediction. The sensor testing method comprises the following steps: generating a multi-dimensional dynamic excitation sequence containing target gas, various interference gases, temperature and humidity information; synchronously acquiring response data of the sensor in a dynamically changing test environment; a corrected blind source separation and nonlinear regression hybrid model is adopted to decouple the collected response data, and independent response components caused by target gas, interferents and environmental factors are separated out; extracting a group of dynamic key performance indexes (d-KPI s) based on the decoupled data; and predicting the long-term attenuation trend of the dynamic key performance index through a time sequence convolutional network and an accelerated aging prediction model of an attention mechanism. According to the method, comprehensive and accurate characterization of the performance of the gas sensor and life prediction can be realized, and the screening efficiency and the quality control level of a high-end sensor are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of electronic component technology, and more specifically, relates to a sensor testing method based on dynamic environment simulation and performance degradation prediction. Background Technology

[0002] In the mass production of gas sensors, testing is a core process for ensuring product quality and performance grading. Automated testing systems are widely used in current technologies. These systems typically consist of a gas path control unit, a test chamber, a data acquisition unit, and a main control computer, capable of automatically testing the sensor's basic performance indicators (such as sensitivity, response recovery time, and repeatability). However, existing testing methods have the following significant limitations: static testing cannot reflect dynamic performance, cross-sensitivity assessment is insufficient, long-term performance prediction capabilities are lacking, and there is a contradiction between testing efficiency and dimensionality.

[0003] Therefore, there is an urgent need to develop a novel gas sensor testing method that can not only efficiently and comprehensively characterize the multi-dimensional dynamic performance of sensors under near-real-world application conditions, but also scientifically predict their long-term performance degradation trends, thereby enabling more precise quality control and value grading of sensor products. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a sensor testing method based on dynamic environment simulation and performance degradation prediction. This method aims to solve the technical problems of existing testing methods being unable to effectively simulate real dynamic environments, accurately assessing cross-sensitivity, lacking the ability to predict long-term performance stability, and balancing testing efficiency with testing dimensions.

[0005] This method includes the following core steps:

[0006] 1. Multidimensional Dynamic Excitation Sequence Generation: First, the algorithm generates a complex, continuously varying test excitation sequence over time. This sequence is not a simple step signal, but simultaneously incorporates dynamic changes in the target gas concentration, the concentration of one or more typical interfering gases, ambient temperature, and ambient humidity. These variation curves are designed to fully excite the sensor's dynamic response characteristics with non-periodic, wide-spectrum signals, thereby comprehensively probing the sensor's performance under different rates of change and combinations of conditions within a compact testing process.

[0007] 2. Synchronous Acquisition and Preprocessing of Sensor Response: The sensor under test is placed in a test system capable of accurately reproducing the aforementioned dynamic excitation sequence. The system synchronously records the sensor's raw output signal (such as resistance, voltage, or current) and the actual measured values ​​of various parameters in the environment. The acquired raw data undergoes a preprocessing module to remove conventional noise and baseline drift.

[0008] 3. Dynamic Response Decoupling Based on Hybrid Model: This is one of the key technologies of this invention. Since the sensor output is the result of the nonlinear coupling of multiple factors such as the target gas, interfering gas, temperature, and humidity, this invention employs a "Modified Blind Source Separation and Nonlinear Regression Hybrid Model" (M-BSS-NR). This model can mathematically "decouple" the independent response components corresponding to each excitation source (target gas, interfering gas, etc.) from the single sensor output signal, and quantify the nonlinear effects of temperature and humidity on sensitivity and baseline.

[0009] 4. Extraction of Dynamic Key Performance Indicators (d-KPIs): Based on the decoupled pure response signal, this invention defines and extracts a series of novel dynamic key performance indicators (d-KPIs). These indicators surpass traditional static parameters, such as: the "dynamic sensitivity decay factor" to characterize the rate of change of sensitivity with gas concentration; the "interference suppression ratio transient spectrum" to quantify the ability to resolve target gas signals under dynamic temperature and humidity interference; and the "response tracking fidelity" to evaluate how closely the sensor output tracks changes in target gas concentration.

[0010] 5. Performance Degradation Prediction Based on Accelerated Aging: After completing dynamic performance characterization, the sensor undergoes a short-term accelerated aging process. Then, the d-KPI sequence data measured before and after aging are input into a pre-trained Temporal Convolutional Attention Network (T-CAN) prediction model trained with extensive experimental data. This model can learn and capture the evolution of d-KPIs during the aging process, and based on this, predict the complete performance degradation curve of the sensor under future normal use conditions.

[0011] Compared with the prior art, the present invention has the following significant advantages:

[0012] 1. Authenticity and comprehensiveness of the test: By simulating a dynamic, multi-component environment, the test results of this method can more realistically reflect the performance of the sensor in actual application scenarios, overcoming the limitations of traditional static testing.

[0013] 2. Precise cross-sensitivity quantification: The innovative dynamic decoupling algorithm can accurately separate and evaluate cross-sensitivity effects, providing accurate data support for sensor selectivity optimization and subsequent software compensation algorithm development.

[0014] 3. Achieves efficient prediction of long-term stability: By combining accelerated aging and advanced deep learning prediction models, this invention can complete the scientific prediction of the sensor's performance life of up to several years in the production process (within hours), solving the pain point of the traditional aging test cycle being too long.

[0015] It greatly improves testing efficiency and information depth: In a continuous dynamic test, much more performance information can be obtained than with traditional methods, achieving simultaneous improvement in testing efficiency and information depth, which is particularly suitable for large-scale, high-quality sensor production lines. Attached Figure Description

[0016] Figure 1 This is the execution flow of a sensor testing method based on dynamic environment simulation and performance degradation prediction according to the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this invention.

[0018] This invention proposes a sensor testing method based on dynamic environment simulation and performance degradation prediction. The execution process and technical details of this method will be described in detail below.

[0019] The test object is a metal-oxide-semiconductor (MOS) type ethanol (C2H5OH) gas sensor, whose main interfering gases are carbon monoxide (CO) and methane (CH4). The test system hardware platform includes: a high-precision mass flow controller (MFC) array, a gas mixing chamber, a temperature and humidity controllable environmental test chamber, a high-precision data acquisition card, and a central control computer. This hardware platform can generate dynamically changing gas concentration, temperature, and humidity environments according to software instructions.

[0020] Definitions and Terminology Explanation

[0021] Generate_Excitation_Sequence(): A function for generating dynamic excitation sequences.

[0022] Acquire_And_Preprocess_Data(): Data acquisition and preprocessing function.

[0023] Decouple_Response_Signal(): Function to decouple response signals.

[0024] Extract_dKPIs(): Dynamic key performance indicator extraction function.

[0025] Predict_Aging_Trajectory(): A function to predict performance aging trajectories.

[0026] Generate_Performance_Fingerprint(): The function for generating performance fingerprints.

[0027] Chirp signal: A linearly modulated frequency signal, referring to a signal whose frequency changes linearly with time. In this invention, it is used to construct an excitation sequence to detect the sensor response at different rates of change.

[0028] Blind Source Separation (BSS): A signal processing technique designed to recover the original, independent source signals from a mixed signal with little or no prior knowledge of the source signals or the mixing process.

[0029] Temporal Convolutional Network (TCN): A deep learning network architecture suitable for processing time series data. It uses causal convolution to effectively capture long-term dependencies in time series.

[0030] Self-Attention Mechanism: A mechanism in deep learning that allows models to dynamically assign different importance weights to different parts of a sequence when processing sequential data.

[0031] Combination Figure 1 Detailed explanation of the specific algorithm implementation:

[0032] Step 1: Dynamic Test Excitation Sequence Generation (Generate_Excitation_Sequence)

[0033] This step aims to generate an information-rich sequence of environmental changes that can fully stimulate the dynamic characteristics of the sensor. Traditional step signals can only test one frequency point, while this invention uses a multi-frequency superimposed chirp signal to scan the sensor's response at different rates of change in a single test.

[0034] Step 1.1: Define the basic chirp signal unit

[0035] A basic chirp signal unit is defined as:

[0036]

[0037] The instantaneous value of the basic Chirp signal is equal to the amplitude multiplied by a sine function, where the phase of the sine function is the sum of twice pi multiplied by the initial frequency multiplied by time, plus half the constant of the rate of change of frequency multiplied by the square of time, plus the initial phase.

[0038] Variable definition and value acquisition:

[0039] t: Time variable, in seconds (s).

[0040] S ch irp (t): Chirp signal value at time t.

[0041] A: Signal Amplitude. For gas concentration, the unit is ppm (parts per million); for temperature, the unit is degrees Celsius (°C); for humidity, the unit is percentage of relative humidity (%RH). Its range depends on the range of the sensor under test and the application scenario. For example, the amplitude for ethanol concentration can be 50 ppm.

[0042] f0: Initial Frequency, measured in Hertz (Hz), represents the rate of change of the excitation signal at the start. For example, 0.001 Hz can be used to indicate the start of a slow change.

[0043] k: Chirp Rate, a constant representing the rate of change of frequency, measured in Hz / s. A value >0 indicates frequency increase, and <0 indicates frequency decrease. Controlling this value allows control over the range of the excitation signal's rate of change. For example, this value can be 0.0005 Hz / s.

[0044] Initial phase, in radians. It can be a random value between 0 and 2 / π to increase the randomness of the sequence.

[0045] Step 1.2: Construct a multidimensional composite excitation sequence

[0046] By linearly superimposing multiple parameters Different basic chirp signal units are used, along with a DC bias (baseline), to generate the final excitation sequence of target gas, interfering gas, temperature, and humidity.

[0047]

[0048] The instantaneous concentration of the target gas is equal to the baseline value of the target gas concentration plus the sum of multiple basic Chirp signal units of the target gas, where each unit is its amplitude multiplied by the corresponding sine function.

[0049] Variable definition and value acquisition:

[0050] C target (t): The target gas (ethanol) concentration set at time t, in ppm.

[0051] B E Baseline value of target gas concentration, i.e., the average center value of the concentration. For example, a value of 100 ppm.

[0052] N E : The number of Chirp units used to superimpose the target gas signal. For example, a value of 3.

[0053] A E,i ,f E0,i ,k E,i , The parameters of the i-th Chirp unit used to superimpose the target gas signal. The combination of these parameters is designed to ensure that the final sequence has a wide and uniform spectral coverage.

[0054] Step 2: Data Acquisition and Preprocessing

[0055] Step 2.1: Synchronous Data Acquisition

[0056] At a higher sampling frequency (e.g., f) s =10Hz), synchronously acquire the sensor's resistance output value R raw (t) and actual environmental parameters measured by a high-precision reference sensor inside the test chamber: the concentration sequence of the interfering gas CO C CO (t), Concentration sequence of interfering gas CH4 C CH4 (t), ambient temperature sequence T env (t) and ambient humidity sequence H env (t). The total duration T of the entire test sequence. total (For example, 3600 seconds) is determined by the time required to cover the lowest frequency.

[0057] Step 2.2: Signal Denoising Based on Wavelet Transform

[0058] Due to the sensor's original signal R raw (t) contains thermal noise and electronic noise, which are denoised using wavelet transform.

[0059] 1. For R raw (t) Perform multi-level Discrete Wavelet Transform (DWT) to obtain the wavelet coefficients of each level.

[0060] 2. Apply a soft thresholding function to the detail coefficients in the high-frequency range. The threshold can be determined based on noise level estimation (such as MedianAbsoluteDeviation).

[0061] 3. Perform inverse discrete wavelet transform (IDWT) to obtain the denoised signal R. denoised (t).

[0062] Compared to traditional mean filtering, this method can better preserve the abrupt changes in the signal.

[0063] Step 2.3: Baseline Drift Correction

[0064] Sensor signals may exhibit baseline drift due to aging or slow temperature changes. An asymmetric least-squares (ALS) smoothing algorithm is employed to fit and remove the baseline. This algorithm introduces an asymmetric weight into the least-squares fitting, enabling it to effectively fit the lower envelope (i.e., the baseline) of the signal while ignoring the peak components. The resulting preprocessed signal R is then obtained. preprocessed (t).

[0065] Step 3: Dynamic Decoupling of Cross-Sensitivity (Decouple_Response_Signal)

[0066] This step aims to extract from the mixed response R preprocessed The independent contributions of each factor are separated in (t).

[0067] Step 3.1: Construct the modified blind source separation (M-BSS) model

[0068] We assume that the sensor's response can be modeled as a superposition of a linear mixing process and a nonlinear modulation process.

[0069]

[0070] The change in sensor resistance equals the nonlinear gain function multiplied by the linear weighted sum of the responses of each gas source, plus a noise term. The nonlinear gain function is a function of the actual ambient temperature and humidity.

[0071] Variable definition and value acquisition:

[0072] ΔR(t): The change in sensor resistance relative to the reference resistance R0. R0 is the sensor's resistance value under clean air and standard temperature and humidity conditions, which can be pre-calibrated.

[0073] S j (t): The "pure" response of the j-th gas source, which we assume here is similar to the actual gas concentration C'. j(t) is directly proportional.

[0074] a j : Mixing coefficient, representing the sensor's basic sensitivity to the j-th gas source.

[0075] G nonlin (T' env (t)'H' env (t)): Nonlinear gain modulation function, characterizing the effect of temperature and humidity on overall sensitivity.

[0076] ∈(t): Model residuals.

[0077] Step 3.2: Model Solving and Decoupling

[0078] The solution process is carried out in two iterative steps:

[0079] 1. Linear BSS separation: Fixed G nonlin =1, using the classic Independent Component Analysis (ICA) algorithm, with C' target (t),C' CO (t),C' CH4 Using ΔR(t) as a reference signal, the initial mixing coefficient a is estimated from ΔR(t). j and source response S j (t).

[0080] 2. Nonlinear gain fitting: The linear mixture result ∑a obtained in the previous step is used to fit the nonlinear gain. j S j (t) is a known term, and the function G is fitted using the nonlinear least squares method. nonlin (T' env (t),H' env (t)). This function can be a polynomial model or a small neural network. For example, a second-order polynomial model is:

[0081] G nonlin (T, H) = c0 + c1T + c2H + c3T 2 +c4H 2 +c5TH

[0082] The nonlinear gain function equals a constant term plus a first-order term of temperature plus a first-order term of humidity plus a second-order term of temperature plus a second-order term of humidity plus a cross term of temperature and humidity.

[0083] Variable definitions: T and H are temperature and humidity variables, respectively; c0,...,c5 are the coefficients to be fitted.

[0084] Iterate through these two steps until the coefficient a j and ck Convergence. Finally, we obtain the decoupled pure response R of the target gas. target (t)=a E ·S E (t) and each interference term.

[0085] Step 4: Extracting Dynamic Key Performance Indicators (Extract_dKPIs)

[0086] Based on the decoupled pure response R target (t) and actual target gas concentration C' target (t), calculate a series of innovative d-KPIs.

[0087] Indicator 1: Dynamic Sensitivity Attenuation Factor (DSAF)

[0088] Defined as the ratio of dynamic sensitivity to quasi-static sensitivity. Dynamic sensitivity is calculated when the signal changes rapidly, while quasi-static sensitivity is calculated when the signal changes slowly.

[0089] 1. Calculate the rate of change of the target gas concentration.

[0090] 2. The entire testing process is divided into... The size is divided into "dynamic zone" and "quasi-static zone" based on the threshold.

[0091] 3. Calculate the average sensitivity S in these two regions respectively. dyn and S static Sensitivity is defined as S = (ΔR / R0) / ΔC, where ΔC is the change in gas concentration.

[0092] 4. Calculate DSAF:

[0093]

[0094] The dynamic sensitivity attenuation factor is equal to the dynamic sensitivity divided by the quasi-static sensitivity.

[0095] Variable definition: η DSAF This is the dynamic sensitivity attenuation factor, which is dimensionless. The closer its value is to 1, the better the dynamic performance of the sensor.

[0096] Indicator 2: Transient Spectrum of Interference Rejection Ratio (TSIRR)

[0097] This metric measures the selectivity of a sensor under different frequencies of disturbance variation.

[0098] 1. Response R to the decoupled target gas target (t) and total disturbance response R interfere (t)=∑ j∈{CO，CH4} a j S j Perform a short-time Fourier transform (STFT) on (t) to obtain their energy distribution P in the time-frequency domain. target (t, f) and P interfere (t, f).

[0099] 2. Calculate TSIRR:

[0100]

[0101] The interference suppression ratio transient spectrum is equal to 10 multiplied by the logarithm to the base 10, where the logarithm is the time-frequency energy of the target gas response divided by the time-frequency energy of the total interference response.

[0102] Variable definition: ρ TSIRR (t, f) represents the interference rejection ratio at the time-frequency point (t, f), in dB. This index is a two-dimensional matrix that comprehensively describes the variation of selectivity with time and interference frequency.

[0103] Metric 3: Response Tracking Fidelity (RTF)

[0104] The normalized cross-correlation function is used to evaluate the decoupled response R. target (t) and actual concentration C' target (t) Similarity in shape and time delay.

[0105]

[0106] Response tracking fidelity is a function of time delay, and its value is equal to the cross-correlation integral of the normalized target gas response signal and the time-shifted normalized actual target gas concentration signal.

[0107] Variable definition: γ RTF (τ) is the cross-correlation coefficient when the delay is τ. and This indicates that the DC component of the signal has been removed. γ RTF The maximum value of (τ) represents the shape similarity, and the corresponding τ value is the average response delay.

[0108] Step 5: Accelerated Aging Prediction Based on Temporal Convolutional Attention Networks (Predict_Aging_Trajectory)

[0109] Step 5.1: Accelerated aging test

[0110] After the initial testing, the sensor is placed in an accelerated aging environment. Accelerated stress is applied to the sensor under test to simulate performance degradation before long-term prediction. Common stress conditions include high temperature (e.g., 85°C), high humidity (e.g., 85% RH), or high load voltage, for a set time (e.g., 24 hours). After aging, steps one through four are repeated to obtain a set of aged d-KPIs.

[0111] Step 5.2: T-CAN Prediction Model

[0112] The model has been pre-trained on a large dataset of sensor aging data. The input to the model is the d-KPIs vector sequence X = [dKPIs before and after aging]. initial dKPIs post_aging The model's architecture includes:

[0113] 1. A TCN module with multiple residual connections is used to extract deep temporal features from the input sequence.

[0114] 2. A self-attention module for identifying key d-KPIs that play a decisive role in performance degradation or their changing characteristics during the aging process, and assigning them weights.

[0115] 3. A fully connected output layer for predicting d-KPI vectors for multiple future time points (e.g., 1 year, 2 years, ..., 5 years).

[0116] The model's output is a prediction matrix.

[0117] Step 5.3: Generate the decay trajectory

[0118] For each d-KPI in the prediction matrix, an exponential decay curve can be fitted, for example:

[0119] P(t) = P0·e -λt +P ∞

[0120] The future performance parameter value equals the initial performance parameter value multiplied by the natural exponential function (whose exponent is a negative decay constant multiplied by time) plus the stable final value of the performance parameter.

[0121] Variable definition: P(t) is a function of any d-KPI changing with time t (in years); P0, λ, P ∞ These are the decay model parameters fitted for each d-KPI.

[0122] Through the above five steps, the testing method proposed in this invention can provide an unprecedentedly in-depth, dynamic, and forward-looking evaluation of the performance of gas sensors, thereby bringing revolutionary improvements to the production and quality control of high-end electronic components.

[0123] Compared with existing technologies, the testing method proposed in this invention has the following significant advantages:

[0124] 1. Authenticity and comprehensiveness of the test: By simulating a dynamic, multi-component environment, the test results of this method can more realistically reflect the performance of the sensor in actual application scenarios, overcoming the limitations of traditional static testing.

[0125] 2. Precise cross-sensitivity quantification: The innovative dynamic decoupling algorithm can accurately separate and evaluate cross-sensitivity effects, providing accurate data support for sensor selectivity optimization and subsequent software compensation algorithm development.

[0126] 3. Achieves efficient prediction of long-term stability: By combining accelerated aging and advanced deep learning prediction models, this invention can complete the scientific prediction of the sensor's performance life of up to several years in the production process (within hours), solving the pain point of the traditional aging test cycle being too long.

[0127] 4. Significantly improves testing efficiency and information dimension: In a continuous dynamic test, much more performance information can be obtained than with traditional methods, achieving simultaneous improvement in testing efficiency and information dimension, which is particularly suitable for large-scale, high-quality sensor production lines.

[0128] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. The scope of protection of the present invention is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to the present invention within its spirit and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of the present invention.

Claims

1. A sensor testing method based on dynamic environment simulation and performance degradation prediction, characterized in that, Includes the following steps: a. Test sequence generation step: Generate one or more test excitation sequences containing the target gas concentration, at least one interfering gas concentration, ambient temperature and ambient humidity dynamically changing over time; b. Data acquisition steps: Place the gas sensor under test in the test environment, drive the environmental control device according to the test excitation sequence, and simultaneously acquire the output response data of the gas sensor under test; c. Response decoupling step: Process the output response data to separate and quantify the mixed response generated by the combined effects of the target gas, the interfering gas, ambient temperature, and ambient humidity; d. Parameter extraction step: Based on the results of the response decoupling step, calculate and extract one or more sets of dynamic key performance indicators characterizing the dynamic performance of the gas sensor under test; e. Aging prediction step: Input the dynamic key performance indicators into a pre-trained performance degradation prediction model to predict the performance change trajectory of the gas sensor under test within a preset working life.

2. The method according to claim 1, characterized in that, In the step of generating the test sequence, the dynamic change curves of the target gas concentration, interfering gas concentration, ambient temperature and ambient humidity are composite signals formed by the linear superposition of multiple Chirp signals with different frequencies and amplitudes.

3. The method according to claim 1, characterized in that, The response decoupling step employs a modified blind source separation algorithm. This algorithm first uses a linear transformation to initially separate the principal components of the signal associated with each excitation source, and then applies a nonlinear mapping function to compensate and correct the cross-coupling effect of temperature and humidity.

4. The method according to claim 1 or 2, characterized in that, The key dynamic performance indicators include: dynamic sensitivity attenuation factor, interference suppression ratio transient spectrum, and response tracking fidelity.

5. The method according to claim 1, characterized in that, The performance degradation prediction model is a deep learning model that integrates temporal convolutional networks and self-attention mechanisms.

6. The method according to claim 1, characterized in that, Prior to the aging prediction step, there is also an accelerated aging step, which simulates the performance degradation of the gas sensor under test over a long period of time by applying periodic high temperature, high humidity or electrical stress to the gas sensor under test.

Images