Instrument aging dynamic test method based on bidirectional LSTM (Long Short Term Memory) and wavelet noise reduction

Abstract

The invention provides an instrument aging dynamic test method based on bidirectional LSTM and wavelet noise reduction. The method comprises the following steps: S1, collecting a time sequence operation data set of a tested instrument; s2, carrying out wavelet denoising processing on the time sequence operation data set to obtain a denoised data set; s3, generating a standardized feature matrix; s4, constructing a bidirectional LSTM neural network model; s5, processing the standardized feature matrix by adopting a sliding window method to generate a training sample set; s6, dividing the training sample set into a training subset and a verification subset according to a ratio of 7: 3, and training a bidirectional LSTM neural network model; s7, instrument data collected in real time are processed through S2-S3 and then input into the model, and an aging degree predicted value # imgabs0 # is output; S8, the aging degree predicted value # imgabs1 # output in the S7 is compared with a dynamic threshold value theta, and when # imgabs2 # is detected, an early warning signal is triggered. According to the invention, the neural network model is adopted to improve the accuracy and reliability of the test result.

Description

Technical Field

[0001] The present invention relates to the technical field of aging testing of instruments and meters, and particularly to a dynamic testing method for instrument aging based on bidirectional LSTM and wavelet denoising. Background Art

[0002] The aging testing of instruments and meters is a method of accelerating product aging by simulating the actual use environment or applying extreme conditions to detect potential defects, evaluate reliability and stability.

[0003] The main purpose of aging testing is to eliminate early failure products caused by process or material defects, ensure the stable performance of instruments and meters during long-term use, and improve product design or production process through test data.

[0004] Common aging testing methods are as follows:

[0005] Environmental aging testing, including high-temperature aging, low-temperature aging, temperature cycling, and damp-heat aging, etc.;

[0006] Electrical stress aging testing, including continuous power-on aging, switch cycle testing, and overvoltage / overcurrent testing, etc.;

[0007] Mechanical stress aging testing, including vibration testing, shock testing, and long-term operation aging, etc.;

[0008] Comprehensive aging testing, combining the above multiple aging tests.

[0009] The above aging testing methods usually rely on manual and testers' experience for testing. For example, fixed thresholds are used to judge the test results, and the processing of test data is relatively rough, resulting in low accuracy and reliability of test results. Summary of the Invention

[0010] The present invention provides a dynamic testing method for instrument aging based on bidirectional LSTM and wavelet denoising, which uses a neural network model to improve the accuracy and reliability of test results.

[0011] To achieve the above object, the present invention adopts the following technical solutions:

[0012] A dynamic testing method for instrument aging based on bidirectional LSTM and wavelet denoising, comprising:

[0013] S1: Collect the time-series operation data set of the instrument under test, and the time-series operation data set includes a voltage fluctuation value set, a current offset amount set, and a temperature response delay time set;

[0014] S2: Perform wavelet denoising processing on the voltage fluctuation value set, the current offset amount set, and the temperature response delay time set respectively to obtain the corresponding denoised data sets;

[0015] S3: For the denoised dataset, use the normalization formula x'i = (xi - μi) / σi to generate the normalized feature matrix X = [x1, x2, x3], where x'i is the data before normalization, xi is the data after normalization, μi and σi are the statistical mean and standard deviation of the historical data of the same type of instrument respectively, and x1, x2, and x3 are the voltage features, current features, and temperature response features after normalization of the denoised datasets corresponding to the voltage fluctuation value set, current offset amount set, and temperature response delay time set respectively;

[0016] S4: Construct a bidirectional LSTM neural network model, and set the number of nodes in the input layer of the model to 3, which is consistent with the dimension of the normalized feature matrix X;

[0017] S5: Use the sliding window method to process the normalized feature matrix X to generate a training sample set;

[0018] S6: Divide the training sample set into a training subset and a validation subset according to 7:3, and use the Adam optimizer to train the bidirectional LSTM neural network model constructed in S4 with the MAE + MSE composite loss function. Terminate the training when the loss of the validation set has not decreased for 50 consecutive epochs;

[0019] S7: Input the instrument data collected in real time after being processed by S2 - S3 into the model trained in S6, and output the aging degree prediction value

[0020] S8: Compare the aging degree prediction value output by S7 with the dynamic threshold θ. When it triggers an early warning signal, where θ = θbase + Δθ, θbase comes from the mean time between failures parameter in the instrument technical manual, and Δθ is the threshold adjustment amount, is the predicted mean in the last 30 days, and Ytrueavg is the arithmetic mean of the true aging degrees actually detected in the last 30 days.

[0021] In this specification, the dynamic testing method for instrument aging based on bidirectional LSTM and wavelet denoising further includes:

[0022] S9: Perform trend analysis on the sequences continuously output by S7 to generate a visualization report including an aging degree prediction curve and inflection point markers. The inflection points are identified by the second derivative mutation detection algorithm.

[0023] In this specification, the dynamic testing method for instrument aging based on bidirectional LSTM and wavelet denoising further includes:

[0024] S10: Dynamically update the test period according to the value of S7. The formula for the new period is Where T new is the updated test period, T old is the initial value of the current test period, and k is the attenuation coefficient.

[0025] In this specification, the wavelet denoising process in S2 includes:

[0026] Performing 5-layer decomposition on the voltage fluctuation value set using the dB4 wavelet basis and denoising using the sqtwolog threshold rule;

[0027] Performing 6-layer decomposition on the current offset amount set using the sym8 wavelet basis and applying the adaptive threshold λ = σ√(2lnN), where σ is the noise standard deviation and N is the signal length;

[0028] Performing median filtering with a window width of 5 points on the temperature response delay time set first, and then performing 3-layer decomposition and reconstruction using the haar wavelet.

[0029] In this specification, the structure of the bidirectional LSTM neural network model in S4 includes:

[0030] The first hidden layer: a 64-unit bidirectional LSTM, returning the complete time series;

[0031] The second hidden layer: a 32-unit bidirectional LSTM, only outputting the state of the final time step;

[0032] The attention mechanism layer, calculating the time step weight αt = softmax(Wa·ht + ba); softmax is the normalized exponential function, ht is the output of the first hidden layer, Wa is the trainable weight matrix, and ba is the bias term;

[0033] The output layer, using the Sigmoid activation function to map the predicted value to the interval [0, 1]

[0034] In this specification, in S5, the sliding window method is used to process the standardized feature matrix X, with a window length T = 24h and a sliding step ΔT = 1h, generating a training sample set D = {(Xi, Yi)|i = 1,..., N}, where Xi is the time series data intercepted by the i-th sliding window in the standardized feature matrix X, and Yi = α·x1i + β·x2i + γ·x3i, and the weight coefficients α, β, γ are extracted from the historical fault data through the principal component analysis method.

[0035] In this specification, the second derivative mutation detection algorithm is:

[0036] The second derivative Δt = 1min;

[0037] When three consecutive sampling points satisfy It is marked as an inflection point at that time.

[0038] In this specification, the test cycle adjustment strategy in S10 includes:

[0039] T new ∈[0.3T old , 1.5T old ;

[0040] When it is continuously ≥ 0.7 for 5 times, T is forced to be set to new 0.3T old .

[0041] In this specification, the model training in S6 includes:

[0042] The loss function L = 0.7MAE + 0.3MSE,

[0043] The Adam optimizer parameters are β1 = 0.9, β2 = 0.999, and the initial learning rate η = 0.001; the learning rate is reduced to 1 / 10 of the original value when the validation loss stagnates.

[0044] In this specification, S1 specifically includes:

[0045] Placing the instrument under test in a temperature control box to execute a temperature cycle program:

[0046] The starting temperature is -20°C and it is heated to +60°C at a rate of 2°C / min;

[0047] Maintain the high temperature for 30 minutes;

[0048] Cool down back to -20°C at a rate of 1.5°C / min;

[0049] Synchronously collect:

[0050] The sampling rate of the voltage fluctuation value is ≥ 1 kHz;

[0051] The sampling interval of the current offset is ≤ 0.1 s;

[0052] The temperature response delay time τ(t) = t2 - t1, where t1 is the time when the set value of the temperature control box changes, and t2 is the time when the displayed value of the instrument reaches the target value ±1%, and the timestamp alignment accuracy is ≤ 1 ms.

[0053] In summary, the present invention has at least the following beneficial effects:

[0054] The present invention uses a neural network model to improve the accuracy and reliability of test results; it realizes high-precision and adaptive aging state evaluation and closed-loop control. Description of the Drawings

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings.

[0056] Figure 1 It is a schematic diagram of the instrument aging dynamic test method based on bidirectional LSTM and wavelet denoising involved in the present invention.

[0057] Figure 2 It is a schematic diagram of the wavelet denoising process in S2 involved in the present invention.

[0058] Figure 3 It is a schematic diagram of denoising V(t) involved in the present invention.

[0059] Figure 4 It is a schematic diagram of denoising I(t) involved in the present invention.

[0060] Figure 5 It is a schematic diagram of denoising τ(t) involved in the present invention. Specific embodiments

[0061] In the following, only some exemplary embodiments are briefly described. As those skilled in the art can recognize, the described embodiments can be modified in various different ways without departing from the spirit or scope of the embodiments of the present invention. Therefore, the accompanying drawings and the description are considered to be exemplary in nature rather than restrictive.

[0062] The following disclosure provides many different embodiments or examples for implementing different structures of the embodiments of the present invention. To simplify the disclosure of the embodiments of the present invention, the components and settings of specific examples are described below. Of course, they are only examples and are not intended to limit the embodiments of the present invention. In addition, the embodiments of the present invention can repeat reference numerals and / or reference letters in different examples. This repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed.

[0063] The following will describe the embodiments of the present invention in detail with reference to the accompanying drawings.

[0064] As Figure 1 shown, this embodiment provides an instrument aging dynamic test method based on bidirectional LSTM and wavelet denoising, including:

[0065] Step S1: Collect the time-series operation data set of the instrument under test under preset environmental conditions. The time-series operation data set includes a voltage fluctuation value set {V(t)}, a current offset set {I(t)}, and a temperature response delay time set {τ(t)}, where t represents the sampling time point;

[0066] Step S2: Perform wavelet denoising processing on {V(t)}, {I(t)}, and {τ(t)} obtained in Step S1 respectively to obtain the denoised data sets {V'(t)}, {I'(t)}, and {τ'(t)};

[0067] Step S3: Input the denoised data set into the normalization processing module, and generate a normalized feature matrix X = [x1, x2, x3] according to the formula x'i = (xi - μi) / σi, where x1 corresponds to V'(t), x2 corresponds to I'(t), x3 corresponds to τ'(t), and μi and σi are the statistical mean and standard deviation of the historical data of the same type of instrument;

[0068] Step S4: Construct a bidirectional LSTM neural network model, and set the number of nodes in the input layer of the model to 3, which is strictly consistent with the dimension of the normalized feature matrix X;

[0069] Step S5: Process the normalized feature matrix X using the sliding window method, with the window length T = 24h and the sliding step ΔT = 1h, to generate a training sample set D = {(Xi, Yi)|i = 1,..., N}, where Yi = α·x1i + β·x2i + γ·x3i, and the weight coefficients α, β, and γ are extracted from the historical fault data by the principal component analysis method;

[0070] Step S6: Divide the training sample set D into a training subset Dtrain and a validation subset Dval according to 7:3, and use the Adam optimizer to train the neural network constructed in Step S4 with the MAE + MSE composite loss function. Terminate the training when the validation set loss does not decrease for 50 consecutive epochs;

[0071] Step S7: Input the instrument data collected in real time after being processed by Steps S2 - S3 into the trained model, and output the aging degree prediction value

[0072] Step S8: Compare the output in Step S7 with the dynamic threshold θ. When a warning signal is triggered, where θ = θbase + Δθ, θbase comes from the mean time between failures parameter in the instrument technical manual, and Δθ is dynamically adjusted by the online learning module according to the prediction error in the recent 30 days;

[0073] Step S9: The continuously output in Step S7 Sequence input trend analysis module (an existing module can be used, such as Python libraries, R language, ARIMA / SARIMA, Holt-Winters, etc.), generate a visualization report containing the aging degree prediction curve and inflection point markers, and the inflection points are identified by a second derivative mutation detection algorithm;

[0074] Step S10: According to the value in step S7, dynamically update the test cycle, and the formula for the new cycle is where T old is the initial value of the current test cycle, and k = 0.85 is the attenuation coefficient, which can be optimized and determined through Monte Carlo simulation.

[0075] In some embodiments, the tool selection for the visualization report:

[0076] Python ecosystem:

[0077] Matplotlib / Seaborn: Static charts (line charts, bar charts);

[0078] Plotly / Dash: Interactive dynamic reports (support hover, zoom, data export);

[0079] Bokeh: Real-time data stream visualization.

[0080] Front-end framework:

[0081] ECharts (JavaScript): Suitable for Web-side embedding;

[0082] PowerBI / Tableau: Enterprise-level automated reports.

[0083] In some embodiments, the Monte Carlo simulation optimization is determined as follows:

[0084] 1. Define the optimization objective and constraints

[0085] Optimization objective: Determine the optimal value of the attenuation coefficient k to maximize the test efficiency (i.e., shorten the unnecessary test cycle) on the premise that the early warning false negative rate < 5%.

[0086] Constraints:

[0087] False Negative Rate (FNR) < 5%;

[0088] Test cycle adjustment range limit: T new ∈[0.3T old , 1.5T old .

[0089] 2. Monte Carlo Simulation Steps

[0090] Step 1: Parameter Space Definition

[0091] Candidate range of attenuation coefficient k: k ∈ [0.5, 1.2] (uniform distribution, i.e., kcandidate ∼ U(0.5, 1.2)).

[0092] Number of simulations: At least 10,000 random samplings to ensure statistical significance.

[0093] Step 2: Aging Scenario Simulation

[0094] Generate multiple sets of simulated aging data covering different aging modes:

[0095] Rapid aging: Aging degree Rises from 0.3 to 0.9 in a short period (e.g., 10 days);

[0096] Linear aging: Increases at a constant rate (e.g., +0.02 per day);

[0097] Random fluctuation aging: Randomly fluctuates within the normal range (0.2 - 0.6);

[0098] Sudden aging: Suddenly jumps after the stable stage (e.g., from 0.4 to 0.8). Step 3: Simulation Run (Single Simulation Process)

[0099] For each kcandidate, perform the following operations:

[0100] Initialize parameters:

[0101] Initial test period T old = 24h;

[0102] Aging degree sequence Generated according to the preset scenario.

[0103] Period adjustment simulation:

[0104] Calculate T at each time point according to the formula new :

[0105]

[0106] Constraint condition: If T new exceeds [0.3T old , 1.5T old , then take the boundary value.

[0107] False alarm detection:

[0108] Record the number of undetected aging events caused by extended cycles;

[0109] Calculate the false negative rate: FNR = (Number of undetected events / Total number of actual aging events) × 100%;

[0110] Step 4: Performance evaluation metrics

[0111] For each k_candidate, calculate the following metrics:

[0112] False negative rate (FNR): Must satisfy FNR < 5%;

[0113] Mean Test Interval (MTI):

[0114]

[0115] Resource Utilization (RU):

[0116]

[0117] Step 5: Screen the optimal k value

[0118] Filtering constraint: Eliminate all k_candidates with FNR ≥ 5%;

[0119] Multi-objective optimization: Among the remaining candidate values, select k that simultaneously satisfies the following conditions:

[0120] MTI is the smallest (high test frequency, low undetected risk);

[0121] RU is the lowest (less resource consumption).

[0122] Robustness verification: Conduct cross-validation on the optimal k, repeat the simulation using an independent dataset to ensure stability.

[0123] In some embodiments, the process of extracting the weight coefficients α, β, γ by the principal component analysis method (PCA) is as follows:

[0124] 1. Data preparation and standardization;

[0125] Historical failure dataset: Collect the failure data of the same type of instrument to construct the data matrix Z ∈ R m×3 , where: m is the number of failure samples;

[0126] Each row represents a failure sample, including three features:

[0127]

[0128] x1: Voltage fluctuation value, x2: Current offset, x3: Temperature response delay time.

[0129] Data standardization:

[0130] Standardize each feature column to eliminate the difference in dimensions:

[0131]

[0132] where μj and σ j are the mean and standard deviation of the j-th feature respectively.

[0133] 2. Covariance matrix calculation;

[0134] Calculate the covariance matrix Σ ∈ R 3×3 :

[0135]

[0136] The covariance matrix reflects the linear correlation between the three features.

[0137] 3. Eigen Decomposition;

[0138] Perform eigen decomposition on the covariance matrix Σ: Σ = UΛU T ;

[0139] where: U ∈ R 3×3 is the eigenvector matrix, and each column is an eigenvector;

[0140] Λ = diag(λ1, λ2, λ3) is the eigenvalue diagonal matrix, and λ1 ≥ λ2 ≥ λ3 ≥ 0.

[0141] The direction of the first principal component: The eigenvector u1 = [u 11 , u 12 , u 13 corresponding to the largest eigenvalue λ1 T .

[0142] 4. Extract the weight coefficients α, β, γ;

[0143] Direct assignment method:

[0144] Normalize the eigenvector of the first principal component as the weight coefficient:

[0145]

[0146] Ensure that α 2 + β 2 + γ 2 = 1.

[0147] Variance contribution rate adjustment method (optional):

[0148] If the weights are required to reflect the actual explanatory power of the principal components, they can be adjusted according to the eigenvalue ratio:

[0149]

[0150] This method makes the weight coefficients proportional to the variance contribution rates of the principal components.

[0151] In some embodiments, step S1 specifically includes:

[0152] Preset environmental conditions:

[0153] Connect the instrument under test (digital multimeter 34401A) to the standard source generator (FLUKE 5522A);

[0154] The temperature control box executes a temperature cycle program: the starting temperature θstart = -20 °C, the heating rate r = 2 °C / min to θmax = +60 °C, hold for t = 30 min and then cool back to θstart at a rate of r = 1.5 °C / min;

[0155] Synchronous acquisition:

[0156] V(t) is sampled at f = 1 kHz through the NIPXIe-4300 data acquisition card;

[0157] I(t) is sampled by the shunt method at Δt = 0.1 s;

[0158] τ(t) = t2 - t1, where t1 is the time when the set value of the temperature control box changes, and t2 is the time when the displayed value of the instrument reaches the target value ±1%, and the timestamp alignment accuracy ≤ 1 ms.

[0159] In some embodiments, the wavelet denoising process of step S2 includes:

[0160] For V(t): Perform L = 5-layer decomposition using the dB4 wavelet basis function, and the threshold rule is sqtwolog (λ = σ√(2lnN));

[0161] For I(t): Perform L = 6-layer decomposition using the sym8 wavelet basis function, with an adaptive threshold λ = σ√(2lnN), and σ = median(|D1|) / 0.6745;

[0162] For τ(t): First perform median filtering with a window width W = 5, and then perform L = 3-layer decomposition and reconstruction using the haar wavelet.

[0163] In some embodiments, in step S3:

[0164] μi and σi are from the factory inspection database of the same type of instrument, and the data volume ≥ 100 units (each instrument contains at least 100 hours of data);

[0165] Update the parameters every time m = 50 new data are added:

[0166]

[0167] where n is the original data volume and m is the newly added data volume.

[0168] In some embodiments, the bidirectional LSTM neural network model in step S4 includes:

[0169] Input layer: 3 nodes receive

[0170] First hidden layer: 64-unit bidirectional LSTM (return sequence);

[0171] Second hidden layer: 32-unit bidirectional LSTM (return final state);

[0172] Attention layer: weight αt = softmax(Wa·ht + ba), where is the output of the first hidden layer;

[0173] Output layer: output with Sigmoid activation function

[0174] In some embodiments, in step S5:

[0175] Sliding window parameters: T = 24h corresponds to N = 1440 sampling points (Δt = 1min), ΔT = 1h corresponds to ΔN = 60 points;

[0176] Weight coefficients α = 0.5, β = 0.3, γ = 0.2 are obtained by performing eigen decomposition on the covariance matrix through principal component analysis.

[0177] In some embodiments, the model training in step S6 includes:

[0178] Loss function L = 0.7MAE + 0.3MSE,

[0179] Adam optimizer parameters β1 = 0.9, β2 = 0.999, initial learning rate η = 0.001;

[0180] After training is completed, the model file is encrypted and stored using AES-256.

[0181] In some embodiments, the calculation rule of the dynamic threshold θ in step S8 is:

[0182] θbase = 1 - exp(-Tobs / MTBF), where Tobs = 72h and MTBF = 5000h;

[0183] is the predicted mean for the last 30 days;

[0184] θ ∈ [0.6, 0.9], and the boundary values are taken when it exceeds.

[0185] In some embodiments, the inflection point detection algorithm in step S9 is:

[0186] Second derivative Δt = 1min;

[0187] When three consecutive sampling points satisfy it is marked as an inflection point.

[0188] In some embodiments, the test cycle adjustment strategy in step S10 includes:

[0189] T new ∈ [0.3T old , 1.5T old ;

[0190] When it is ≥ 0.7 for five consecutive times, forcefully set T new = 0.3T old ;

[0191] Send the cycle parameter to the temperature control box register address 0003H through the Modbus TCP protocol.

[0192] In a specific embodiment:

[0193] Step S1: Collect timing operation data;

[0194] Connect the instrument under test (such as digital multimeter 34401A) to the standard source generator (model FLUKE 5522A);

[0195] Place the instrument under test in a temperature control box (such as ESPEC PH - 032), and set the temperature cycle program:

[0196] Initial temperature: -20°C;

[0197] Heat up to +60°C (at a rate of 2°C / min);

[0198] Maintain the high temperature for 30 minutes;

[0199] Cool down back to -20°C (at a rate of 1.5°C / min);

[0200] Synchronously start the data acquisition device:

[0201] Voltage fluctuation value V(t): Recorded by a high-precision data acquisition card (NI PXIe-4300) or a high-precision voltage sensor (such as Keysight 34470A) at a sampling rate of 1 kHz;

[0202] Current offset I(t): Measured by the shunt method or a Hall current sensor (such as LEM LAH 50-P), with a sampling interval of 0.1 s;

[0203] Temperature response delay τ(t): Record the time t1 when the set value of the temperature control box changes and the time t2 when the meter display value reaches the target value within the range of ±1%, and calculate τ(t) = t2 - t1; or record the temperature change curve through a thermocouple (such as K-type) and calculate τ(t) = t2 - t1;

[0204] Generate the original data set:

[0205] The storage format is a time series {V(t), I(t), τ(t)}, and the time stamp alignment accuracy is ≤1 ms;

[0206] Step S2: Wavelet denoising processing;

[0207] Denoise V(t):

[0208] a. Perform 5-layer decomposition using the dB4 wavelet basis function;

[0209] b. Use the default threshold rule ('sqtwolog') to remove high-frequency noise;

[0210] c. Reconstruct the signal to obtain V'(t);

[0211] Denoise I(t):

[0212] a. Perform 6-layer decomposition using the sym8 wavelet basis function;

[0213] b. Apply the adaptive threshold: λ = σ√(2lnN), where σ is the noise standard deviation and N is the signal length;

[0214] c. Reconstruct the signal to obtain I'(t);

[0215] Denoise τ(t):

[0216] a. First perform median filtering (window width = 5 sampling points);

[0217] b. Then perform 3-layer decomposition and reconstruction using the haar wavelet;

[0218] c. Output τ'(t);

[0219] Output denoised dataset: {V'(t), I'(t), τ'(t)};

[0220] Step S3: Standardization processing;

[0221] Obtain historical data statistical parameters of instruments of the same model:

[0222] Mean values μv, μi, μτ;

[0223] Standard deviations σv, σi, στ;

[0224] (Source: Instrument factory inspection database, containing at least 100 data of the same model);

[0225] Perform standardization on the denoised data:

[0226] x1 = (V'(t) - μv) / σv;

[0227] x2 = (I'(t) - μi) / σi;

[0228] x3 = (τ'(t) - μτ) / στ;

[0229] Generate a standardized feature matrix X = [x1, x2, x3], with a dimension of N×3 (N is the total number of sampling points); Step S4: Neural network construction;

[0230] Build a bidirectional LSTM network structure:

[0231] Input layer: 3 nodes, receiving the standardized feature matrix X;

[0232] The first hidden layer: Bidirectional LSTM (64 units, return_sequences = True);

[0233] The second hidden layer: Bidirectional LSTM (32 units);

[0234] Attention layer: Calculate the time step weights, with the formula at = softmax(Wa·ht + ba);

[0235] Output layer: Fully connected layer (1 node, Sigmoid activation);

[0236] Parameter initialization:

[0237] Weights: Initialized with He normal distribution;

[0238] Biases: Initialized to 0.01;

[0239] Step S5: Training sample generation;

[0240] Sliding window parameter setting:

[0241] The window length T = 24h (corresponding to 1440 sampling points, with an interval of 1min);

[0242] The sliding step size ΔT = 1h (60 sampling points);

[0243] Perform window splitting on the standardized matrix X:

[0244] Generate the input sample Xi = X[i:i+T,:], i = 0, 60, 120,...

[0245] Calculate the supervised label Yi:

[0246] Obtain the principal component analysis weight coefficients (α = 0.5, β = 0.3, γ = 0.2);

[0247] Yi = 0.5x1 + 0.3x2 + 0.2x3 (calculated for the last time step within the window);

[0248] Generate the training sample set D = {(X1, Y1),..., (XN, YN)};

[0249] Step S6: Model training;

[0250] Data partitioning:

[0251] Training set Dtrain: The first 70% of the samples;

[0252] Validation set Dval: The last 30% of the samples;

[0253] Training configuration:

[0254] Optimizer: Adam (β1 = 0.9, β2 = 0.999);

[0255] Loss function: L = 0.7MAE + 0.3MSE;

[0256] Batch size: 32;

[0257] Training process:

[0258] Maximum number of iterations: 500 epochs;

[0259] Early stopping condition: The validation loss has not decreased for 50 consecutive epochs;

[0260] Learning rate scheduling: Reduce to 1 / 10 of the original value when the loss stagnates;

[0261] Output the trained model parameter file;

[0262] Step S7: Real-time prediction;

[0263] Data flow processing:

[0264] Create a circular buffer on the edge device (NVIDIA Jetson Nano);

[0265] Continuously receive the normalized data stream output in step S3;

[0266] When the buffer accumulates 24 hours of data, trigger the prediction;

[0267] Model inference:

[0268] Load the trained TensorRT engine;

[0269] Adjust the input data dimension to (1, 1440, 3);

[0270] Output the predicted value (Keep 4 decimal places);

[0271] Store the prediction result in the time series database (InfluxDB);

[0272] Step S8: Aging warning;

[0273] Dynamic threshold calculation:

[0274] The basic threshold θ_base = 0.75 (converted from the instrument manual MTBF = 5000h);

[0275] Adjustment amount (Data for the last 30 days);

[0276] The final threshold θ = θ_base + Δθ (limited between 0.6 and 0.9);

[0277] Warning logic:

[0278] When Generate a level 3 warning signal:

[0279] Yellow warning, record the log;

[0280] Orange warning, trigger the self-check program;

[0281] Red warning, force shutdown;

[0282] Step S9: Visual report generation;

[0283] Data input:

[0284] Retrieve a 72-hour continuous sequence from the time series database;

[0285] Inflection point detection:

[0286] Calculate the second derivative:

[0287] Mark the inflection point: When and it lasts for more than 3 sampling points;

[0288] Generate the report content:

[0289] The line chart shows the curve changing with time;

[0290] Highlight the time period exceeding θ in red;

[0291] Statistically count the number of warnings and the maximum aging degree in a table;

[0292] Step S10: Adjust the test period;

[0293] Execute the formula update:

[0294] The current period T old Read from the system configuration file;

[0295] Calculate the new period:

[0296] Constraint condition: T new ∈[0.3T old , 1.5T old ;

[0297] Update strategy:

[0298] When it is ≥0.7 for 5 consecutive times, immediately set T new = 0.3T old ;

[0299] In other cases, perform the period update once every 24 hours;

[0300] Feedback control:

[0301] Write T new into the configuration register of the test controller;

[0302] Send the "period reset" instruction to the temperature control box through the Modbus TCP protocol.

[0303] The above-described embodiments are used to illustrate the present invention, not to limit the present invention. Therefore, changes in the example values or replacement of equivalent elements still belong to the scope of the present invention.

[0304] From the above detailed description, those of ordinary skill in the art can clearly understand that the present invention can indeed achieve the foregoing objectives and has actually met the requirements of the patent law.

[0305] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications to these embodiments once they learn the basic creative concept. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications that fall within the scope of the present invention. The above description is only the preferred embodiments of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be included within the protection scope of the present invention.

[0306] It should be noted that the above description of the process is only for illustration and explanation, and does not limit the scope of application of this specification. For those skilled in the art, various corrections and changes can be made to the process under the guidance of this specification. However, these corrections and changes are still within the scope of this specification.

[0307] The basic concept has been described above. Obviously, for those of ordinary skill in the art after reading this application, the above invention disclosure is only for illustration and does not constitute a limitation to this application. Although not explicitly stated here, those of ordinary skill in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, so such modifications, improvements, and corrections still belong to the spirit and scope of the exemplary embodiments of this application.

[0308] At the same time, this application uses specific terms to describe the embodiments of this application. For example, "one embodiment", "an embodiment", and / or "some embodiments" mean a certain feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that the "one embodiment" or "an embodiment" or "an alternative embodiment" mentioned twice or more at different positions in this specification does not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application can be appropriately combined.

[0309] In addition, those of ordinary skill in the art can understand that various aspects of this application can be illustrated and described by several patentable types or situations, including any new and useful process, machine, product, or combination of substances, or any new and useful improvement thereof. Therefore, various aspects of this application can be implemented entirely by hardware, can be implemented entirely by software (including firmware, resident software, microcode, etc.), or can be implemented by a combination of hardware and software. The above hardware or software can both be referred to as "unit", "module", or "system". In addition, various aspects of this application can take the form of a computer program product embodied in one or more computer-readable media, in which computer-readable program code is included.

[0310] The computer program code required for the operations of various parts of this application can be written in any one or more programming languages, including object-oriented programming languages such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, Python, etc., conventional procedural programming languages such as C programming language, VisualBasic, Fortran2103, Perl, COBOL2102, PHP, ABAP, dynamic programming languages such as Python, Ruby, and Groovy, or other programming languages, etc. This program code can run entirely on the user's computer, or run on the user's computer as an independent software package, or partially run on the user's computer and partially run on a remote computer, or run entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer through any network form, such as a local area network (LAN) or a wide area network (WAN), or connected to an external computer (for example, through the Internet), or in a cloud computing environment, or used as a service such as software as a service (SaaS).

[0311] In addition, unless clearly stated in the claims, the order of the processing elements and sequences, the use of numbers and letters, or the use of other names in this application are not used to limit the order of the processes and methods of this application. Although some currently considered useful embodiments of the invention are discussed through various examples in the above disclosure, it should be understood that such details only serve the purpose of illustration, and the appended claims are not limited to the disclosed embodiments. On the contrary, the claims are intended to cover all modifications and equivalent combinations that conform to the essence and scope of the embodiments of this application. For example, although the implementation of the above various components can be embodied in a hardware device, it can also be implemented as a pure software solution, for example, installed on an existing server or mobile device.

[0312] Similarly, it should be noted that, in order to simplify the description of this application disclosure and thus help the understanding of one or more embodiments of the invention, in the description of the embodiments of this application above, sometimes multiple features are merged into one embodiment, drawing, or description thereof. However, this method of this application should not be construed as reflecting an intention that the claimed subject matter requires more features than those explicitly recited in each claim. On the contrary, the subject matter of the invention should have fewer features than the above single embodiment.

Claims

1. A dynamic testing method for instrument aging based on bidirectional LSTM and wavelet denoising, characterized in that, Including: S1: Collect the time-series operation data set of the instrument under test, where the time-series operation data set includes a voltage fluctuation value set, a current offset amount set, and a temperature response delay time set; S2: Perform wavelet denoising processing on the voltage fluctuation value set, the current offset amount set, and the temperature response delay time set respectively to obtain the corresponding denoised data sets; S3: For the denoised data sets, use the normalization formula x'i = (xi - μi) / σi to generate the normalized feature matrix X = [x1, x2, x3], where x'i is the data before normalization, xi is the data after normalization, μi and σi are the statistical mean and standard deviation of the historical data of the same type of instrument respectively, and x1, x2, and x3 are the voltage features, current features, and temperature response features after normalization of the corresponding denoised data sets of the voltage fluctuation value set, the current offset amount set, and the temperature response delay time set; S4: Construct a bidirectional LSTM neural network model, and set the number of nodes in the input layer of the model to 3, which is consistent with the dimension of the normalized feature matrix X; S5: Use the sliding window method to process the normalized feature matrix X to generate a training sample set; S6: Divide the training sample set into a training subset and a validation subset according to 7:3, and use the Adam optimizer to train the bidirectional LSTM neural network model constructed in S4 with the MAE + MSE composite loss function, and terminate the training when the loss of the validation set does not decrease for 50 consecutive epochs; S7: Input the instrument data collected in real time into the model trained in S6 after being processed by S2 - S3, and output the predicted aging degree value S8: Compare the predicted aging degree value output by S7 with the dynamic threshold θ. When it triggers a warning signal, where θ = θbase + Δθ, θbase is from the mean time between failures parameter in the instrument technical manual, and Δθ is the threshold adjustment amount, is the predicted mean value for the most recent 30 days, Ytrueavg is the arithmetic mean of the true values of the aging degree actually detected in the recent 30 days.

2. The dynamic testing method for instrument aging based on bidirectional LSTM and wavelet denoising according to claim 1, characterized in that, Also including: S9: Perform trend analysis on the sequence continuously output by S7, generate a visualization report including an aging degree prediction curve and inflection point markers, and identify the inflection points through a second derivative mutation detection algorithm.

3. The dynamic test method for instrument aging based on bidirectional LSTM and wavelet denoising according to claim 1, characterized in that Also including: S10: Dynamically update the test period according to the value in S7, and the calculation formula for the new period is where Tis new the updated test period, T old is the initial value of the current test period, and k is the attenuation coefficient.

4. The dynamic testing method for instrument aging based on bidirectional LSTM and wavelet denoising according to claim 1, characterized in that The wavelet denoising processing in S2 includes: Perform 5-layer decomposition on the voltage fluctuation value set using the dB4 wavelet basis and denoise using the sqtwolog threshold rule; Perform 6-layer decomposition on the current offset amount set using the sym8 wavelet basis and apply the adaptive threshold λ = σ√(2lnN), where σ is the noise standard deviation and N is the signal length; For the temperature response delay time set, first perform median filtering with a window width of 5 points, and then perform 3-layer decomposition and reconstruction using the haar wavelet.

5. The dynamic testing method for instrument aging based on bidirectional LSTM and wavelet denoising according to claim 1, wherein The structure of the bidirectional LSTM neural network model in S4 includes: The first hidden layer: a 64-unit bidirectional LSTM, returning the complete time series; The second hidden layer: a 32-unit bidirectional LSTM, only outputting the state of the final time step; The attention mechanism layer, calculating the time step weight αt = softmax(Wa·ht + ba); softmax is the normalized exponential function, ht is the output of the first hidden layer, Wa is the trainable weight matrix, and ba is the bias term; The output layer, using the Sigmoid activation function to map the predicted value to the interval [0, 1].

6. The dynamic test method for instrument aging based on bidirectional LSTM and wavelet denoising according to claim 1, characterized in that In S5, use the sliding window method to process the normalized feature matrix X, with a window length T = 24h and a sliding step ΔT = 1h, to generate a training sample set D = {(Xi, Yi)|i = 1,..., N}, where Xi is the time series data intercepted by the i-th sliding window in the normalized feature matrix X, and Yi = α·x1i + β·x2i + γ·x3i, and the weight coefficients α, β, and γ are extracted from the historical fault data through the principal component analysis method.

7. The dynamic test method for instrument aging based on bidirectional LSTM and wavelet noise reduction according to claim 2, characterized in that, The second-order derivative mutation detection algorithm is: Second derivative Δt = 1 min; When three consecutive sampling points satisfy it is marked as an inflection point.

8. The dynamic testing method for instrument aging based on bidirectional LSTM and wavelet noise reduction according to claim 1, characterized in that, The test cycle adjustment strategy in S10 includes: T new ∈ [0.3T old , 1.5T old ; When When it is ≥ 0.7 for 5 consecutive times, forcefully set T new = 0.3T old .

9. The dynamic test method for instrument aging based on bidirectional LSTM and wavelet noise reduction according to claim 1, characterized in that The model training in S6 includes: The loss function \(L = 0.7MAE+0.3MSE\). The parameters of the Adam optimizer are \(\beta_1 = 0.9\), \(\beta_2 = 0.999\), and the initial learning rate \(\eta = 0.001\); the learning rate is reduced to 1 / 10 of the original value when the validation loss stagnates.

10. The dynamic test method for instrument aging based on bidirectional LSTM and wavelet noise reduction according to claim 1, wherein S1 specifically includes: Placing the instrument under test in a temperature control box to execute a temperature cycling program: The starting temperature is -20°C, and it is heated to +60°C at a rate of 2°C / min; Maintaining the high temperature for 30 min; Cooling back to -20°C at a rate of 1.5°C / min; Synchronously collecting: The sampling rate of the voltage fluctuation value is ≥1 kHz; The sampling interval of the current offset is ≤0.1 s; The temperature response delay time τ(t) = t2 - t1, where t1 is the time when the set value of the temperature control box changes, and t2 is the time when the displayed value of the instrument reaches ±1% of the target value, and the timestamp alignment accuracy is ≤1 ms.

Images