An environmental testing abnormality interruption early warning and overtest protection method and device

Abstract

This invention relates to a method and apparatus for early warning of abnormal interruptions and protection against overtesting in environmental testing. The method involves acquiring multi-source time-series data of the environmental testing process and control command data synchronized with the multi-source time-series data; preprocessing the multi-source time-series data and constructing a model input sequence by combining the control command data from multiple past time points with the preprocessed multi-source time-series data; inputting the model input sequence into a probabilistic time-series prediction model to predict future multi-step prediction results based on the multi-source time-series data; performing risk calculation to obtain risk indicators; if the risk indicators exceed a risk threshold, performing anomaly judgment and handling operations; otherwise, performing overtesting assessment operations. This invention utilizes multi-source time-series data for probabilistic prediction and trend early warning; it can classify sensor drift and actual anomalies; it also achieves overtesting management; and it improves the safety and effectiveness of testing.

Description

Technical Field

[0001] This invention relates to the field of reliability environmental testing and failure prediction and health management (PHM) technology, and in particular to a method and device for early warning of abnormal interruption and overtest protection in environmental testing. Background Technology

[0002] Environmental tests (such as high and low temperatures, temperature and humidity cycling, vibration, etc.) are typically long in duration and involve frequent changes in operating conditions. Existing test safety strategies often employ "instantaneous over-limit alarms + fixed interlocking thresholds," which presents the following problems: (1) It only alarms when the value exceeds the upper limit, and cannot provide sufficient advance warning for risks that have not yet exceeded the limit but have abnormal trends (such as continuous temperature rise or continuous increase in vibration); (2) Sensors are prone to drift, aging or poor contact during long-term operation, which may lead to false alarms, missed alarms or frequent interruptions of the test. (3) It is difficult to distinguish between equipment-side abnormalities (heater / fan / power supply abnormalities) and abnormalities of the tested object (internal short circuit heating, fire precursors, etc.), resulting in untimely handling or accidental shutdown. (4) When the equivalent stress accumulation is not quantitatively managed, “overtesting” is likely to occur, that is, the test is completed but the prototype suffers hidden damage or even accidental damage due to excessive accumulated stress.

[0003] Therefore, there is an urgent need for an environmental testing abnormality interruption early warning and overtest protection method and device to improve the safety and effectiveness of testing. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and device for early warning of abnormal interruption and protection against overtesting in environmental testing, which can perform probability prediction and trend warning; classify sensor drift and real anomalies; and realize overtesting management.

[0005] The objective of this invention is achieved through the following technical solution: A method for early warning of abnormal interruption and protection against overtesting in environmental testing includes the following steps: S1: Acquire multi-source time-series data of the environmental test process and control command data that is time-synchronized with the multi-source time-series data; S2: Preprocess the multi-source time series data, and construct the model input sequence by combining the control command data from multiple past time points with the preprocessed multi-source time series data; S3: Input the input sequence of the model into the probabilistic time series prediction model to predict the multi-source time series data and obtain the prediction results for the future multiple steps; S4: Risk indicators are obtained by calculating risks based on the actual observation data at the current moment and the results of multi-step predictions in the future from multi-source time series data. S5: If the risk indicator exceeds the risk threshold, an anomaly judgment and handling operation will be performed; otherwise, an overtest evaluation operation will be performed. The anomaly judgment and handling operation is to perform an anomaly source judgment operation; when the anomaly source is a real anomaly, an anomaly interruption warning is output and a protection action is executed; when the anomaly source is sensor drift, a drift alarm is output.

[0006] Furthermore, the multi-source time-series data includes environmental sensor data and experimental equipment operating status data; Step S2 constructs the model input sequence using a sliding window method; The future multi-step prediction results include a predicted mean sequence and a prediction uncertainty parameter; The prediction uncertainty parameter is one or more of variance, covariance matrix, and diagonal approximation of covariance matrix; The probabilistic time-series prediction model is any one or a combination of a long short-term memory network, a gated recurrent unit network, or an attention-based Transformer network. The output probability distribution of the probabilistic time series prediction model is a Gaussian distribution or a Student's t-distribution. The probabilistic time series prediction model is trained by minimizing the negative log-likelihood loss.

[0007] Furthermore, the risk indicator is obtained by fusing the anomaly score with the probability of exceeding limits in multiple future steps; The risk indicator is positively correlated with the anomaly score, and the risk indicator is positively correlated with the probability of exceeding the limit in multiple future steps; The anomaly score is calculated based on the actual observation data at the current moment of multi-source time series data and the results of multi-step prediction in the future. The method for calculating the anomaly score is to first calculate the standardized residuals and then construct the anomaly score based on the standardized residuals. The anomaly score is constructed by aggregating multi-channel residuals in Mahalanobis distance form based on the standardized residuals and performing trend enhancement operations. The future multi-step exceedance probability is obtained by weighting and summing the future exceedance probabilities of each variable calculated based on the future multi-step prediction results.

[0008] Furthermore, the risk threshold is an adaptive threshold; the adaptive threshold is set based on historical normal data and combined with the current working conditions to adjust the risk indicators.

[0009] Furthermore, the adaptive threshold is obtained by estimating the quantiles of the risk indicators in historical normal data by grouping them according to operating conditions, and is updated online as the control command data or test phase changes.

[0010] Furthermore, the protection actions include any one or more of the following: cutting off the power supply to the heating circuit, turning off the main power relay, stopping the vibration output, starting the exhaust or inerting process, locking the test steps and maintaining a safe state, and triggering an audible and visual alarm and remote notification; The anomaly source determination operation constructs a sensor drift index based on multi-source time-series data using a dual-time-scale statistical approach, and obtains a cross-sensor correlation consistency measure through correlation calculation. Then, the anomaly sources are classified according to the sensor drift index and the correlation consistency measure. The sensor drift index is constructed from the difference between the fast timescale smoothed value and the slow timescale smoothed value, and normalized by robust dispersion; the cross-sensor correlation consistency measure is obtained by comparing the correlation structure of multiple past time points with the normal baseline correlation structure.

[0011] Furthermore, the overtest assessment operation is to perform overtest protection actions when the assessment indicators meet the overtest conditions; the assessment indicators are one or both of the remaining safety margin and the expected dose value. The remaining safety margin is calculated by subtracting the cumulative equivalent effect dose from the upper dose limit; The cumulative equivalent stress dose is the value obtained by integrating the nonlinear function of the multi-source time-series data deviating from the reference value over time. The expected dose value is obtained by time integration of a nonlinear function that deviates the predicted mean sequence from the reference value; The overtest protection actions include any one or more of the following: reducing the rate of temperature rise or fall, reducing vibration intensity, shortening the holding time, and entering a conservative operating condition phase or terminating the test and indicating that the dose limit has been reached.

[0012] An environmental test abnormality interruption early warning and over-test protection device, applied to an environmental test abnormality interruption early warning and over-test protection method, includes: The data acquisition module is used to acquire multi-source time-series data of the environmental test process and control command data that is time-synchronized with the multi-source time-series data; The preprocessing and construction module is used to preprocess the multi-source time series data and construct the model input sequence by combining the control command data from multiple past time points with the preprocessed multi-source time series data. The probabilistic time series prediction module is used to input the model input sequence into the probabilistic time series prediction model to predict multi-step prediction results for multi-source time series data. The risk indicator calculation module is used to calculate risk indicators based on the actual observation data at the current moment of multi-source time series data and the results of multi-step future predictions. The exception module is used for exception detection and handling. The pass test evaluation module is used to perform pass test evaluation operations.

[0013] An environmental test abnormality interruption early warning and overtest protection system includes: environmental test equipment, sensor group, test equipment controller, and environmental test abnormality interruption early warning and overtest protection device; the environmental test abnormality interruption early warning and overtest protection device is communicatively connected or electrically interlocked with the test equipment controller.

[0014] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of an environmental test abnormal interruption warning and overtest protection method.

[0015] The beneficial effects of this invention are: It utilizes multi-source time-series data for probability prediction and trend early warning; it can classify sensor drift and real anomalies; it also enables overtest management; and improves the safety and effectiveness of experiments. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0017] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0018] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0019] Example 1: like Figure 1 As shown, a method for early warning of abnormal interruption and protection against overtesting in environmental testing includes the following steps: S1: Acquire multi-source time-series data of the environmental test process and control command data that is time-synchronized with the multi-source time-series data; The multi-source time-series data includes environmental sensor data and experimental equipment operating status data; The multi-source time-series data includes any two or more of the following categories: internal temperature, internal humidity, chamber wall temperature, vibration acceleration, heating power or current, fan speed, door lock status, smoke or combustible gas concentration, internal and external temperature difference, and power supply voltage fluctuation.

[0020] The multi-source time-series data is transmitted or stored in the form of data frames, which include timestamps, channel identifiers, numerical values, units, and quality markers. The quality markers are used to indicate one or more of the following: missing data, over-range data, calibration status, communication anomalies, and sensor self-test failures.

[0021] S2: Preprocess the multi-source time series data, and construct the model input sequence by combining the control command data from multiple past time points with the preprocessed multi-source time series data; Step S2 constructs the model input sequence using a sliding window; Step S2 constructs the sliding window input sequence according to a preset window length, and the sliding window input sequence is the model input sequence.

[0022] The preprocessing includes any one or more of the following: missing value repair, outlier suppression, robust scaling normalization, differential feature construction, time alignment of the control command data, and resampling and synchronization of data at different sampling rates.

[0023] S3: Input the input sequence of the model into the probabilistic time series prediction model to predict the multi-source time series data and obtain the prediction results for the future multiple steps; The future multi-step prediction results include a predicted mean sequence and a prediction uncertainty parameter; The prediction uncertainty parameter is one or more of variance, covariance matrix, and diagonal approximation of covariance matrix.

[0024] The predicted mean sequence is a predicted value for multi-source time series data; The probabilistic time-series prediction model is any one or a combination of a long short-term memory network, a gated recurrent unit network, or an attention-based Transformer network. The output probability distribution of the probabilistic time series prediction model is a Gaussian distribution or a Student's t-distribution to improve robustness to heavy-tailed noise and occasional disturbances.

[0025] Heavy-tailed robust distribution: The output probability distribution is replaced with a student t-distribution instead of a Gaussian distribution to improve robustness to random shocks and non-Gaussian noise.

[0026] The probabilistic time series prediction model is trained by minimizing the negative log-likelihood loss.

[0027] Training is performed by minimizing the negative log-likelihood loss, which enables the simultaneous learning of the predicted mean and the heteroscedastic uncertainty related to the operating conditions. The training samples consist of historical normal test data and data covering typical operating condition transitions.

[0028] S4: Risk indicators are obtained by calculating risks based on the actual observation data at the current moment and the results of multi-step predictions in the future from multi-source time series data. The risk indicator is obtained by fusing anomaly scores with the probability of future multi-step exceedances. Risk indicators are obtained by weighted fusion of anomaly scores and the probability of future multi-step exceedances. The risk indicator is positively correlated with the anomaly score, and the risk indicator is positively correlated with the probability of exceeding the limit in multiple future steps.

[0029] The anomaly score is calculated based on the actual observation data at the current moment of multi-source time series data and the results of multi-step prediction in the future. The method for calculating the anomaly score is to first calculate the standardized residuals and then construct the anomaly score based on the standardized residuals.

[0030] That is, the standardized residual is calculated based on the actual observation data at the current moment of the multi-source time series data and the prediction results of future multiple steps, and then the anomaly score is calculated from the standardized residual.

[0031] The anomaly score is constructed by aggregating multi-channel residuals in Mahalanobis distance form based on the standardized residuals and performing trend enhancement operations.

[0032] Trend enhancement operations are one of the following: exponential weighting, moving average, cumulative, and statistical measures.

[0033] Exponential weighting is exponentially weighted smoothing.

[0034] The future multi-step exceedance probability is obtained by weighted summation of the future exceedance probabilities of each variable calculated based on the future multi-step prediction results. That is, the probability of exceeding the limit in multiple future steps is calculated based on the predicted mean sequence and the predicted uncertainty parameter.

[0035] The probability of exceeding the limit in future multi-step predictions is calculated separately for each controlled variable based on the prediction results in future multi-step predictions, and then weighted and summarized according to the preset variable weights.

[0036] The formula for calculating the probability of exceeding the limit in multiple future steps is as follows: , In the formula, For the probability of exceeding the limit in multiple future steps; For variable index, To observe the number of channels, For variables The weighting coefficients, For variables In the future The probability of exceeding the limit within a step; The weighting coefficients satisfy Furthermore, it can be normalized.

[0037] The For variables In the future The formula for calculating the probability of exceeding the limit within a step is: , In the formula, For the prediction step index, ; Controlled variables The safety limit, To predict the first value of the mean sequence One component; This represents the corresponding standard deviation of the forecast; The cumulative distribution function of the standard normal distribution; This is the chain multiplication operator.

[0038] S5: If the risk indicator exceeds the risk threshold, an anomaly judgment and handling operation will be performed; otherwise, an overtest evaluation operation will be performed. The risk threshold is an adaptive threshold; the adaptive threshold is set based on historical normal data and combined with the current working conditions to set the risk indicators.

[0039] The adaptive threshold is obtained by estimating the quantiles of the risk indicators in historical normal data by grouping them according to operating conditions, and is updated online as the control command data or test phase changes.

[0040] The anomaly judgment and handling operation involves judging the source of the anomaly; when the source of the anomaly is a real anomaly, an anomaly interruption warning is output and a protection action is executed; when the source of the anomaly is sensor drift, a drift alarm is output and a maintenance and handling suggestion is generated.

[0041] True anomalies refer to abnormalities in the experimental process or in the tested object.

[0042] The protective actions include any one or more of the following: cutting off the power supply to the heating circuit, turning off the main power relay, stopping the vibration output, starting the exhaust or inerting process, locking the test procedure and maintaining a safe state, and triggering audible and visual alarms and remote notifications.

[0043] When the source of the anomaly is a real anomaly, the anomaly judgment and processing operation also records the multi-source time-series data fragments, risk indicators and anomaly source classification results when the anomaly interruption warning or protection action is triggered, and generates a traceable event report.

[0044] The event report includes the warning trigger time, the anomaly source classification result, the risk indicator curve, the probability of future multi-step over-limit, the result of the protection action execution, and a data snapshot of the preset time before and after the anomaly.

[0045] The anomaly source determination operation constructs a sensor drift index based on multi-source time-series data using a dual-time-scale statistical approach, and obtains a cross-sensor correlation consistency measure through correlation calculation. Then, the anomaly sources are classified according to the sensor drift index and the correlation consistency measure.

[0046] The sensor drift index is constructed from the difference between the fast timescale smoothing value and the slow timescale smoothing value, and normalized by robust dispersion; the cross-sensor correlation consistency measure is obtained by comparing the correlation structure of multiple past moments with the normal baseline correlation structure, that is, by comparing the current window correlation structure with the normal baseline correlation structure.

[0047] As another design, the cross-sensor correlation consistency metric includes one based on sensor topology. Figure 1 Consistency constraints or consistency criteria based on mutual information are used to improve the reliability of anomaly source classification. This involves representing the sensor coupling relationship as a graph structure and introducing... Figure 1 Consistency constraints, or the use of mutual information consistency criteria to replace the correlation matrix difference measure, can improve the reliability of anomaly source determination.

[0048] The anomaly source determination operation classifies anomaly sources into two types: sensor drift; and genuine anomalies.

[0049] The overtest assessment operation is to perform overtest protection actions when the assessment indicators meet the overtest conditions; the assessment indicators are one or both of the remaining safety margin and the expected dose value.

[0050] The remaining safety margin is calculated by subtracting the cumulative equivalent effect dose from the upper dose limit.

[0051] The cumulative equivalent stress dose is the value obtained by integrating the nonlinear function of the deviation of multi-source time-series data from the reference value over time.

[0052] The pass test condition corresponding to the remaining safety margin is that the remaining safety margin is lower than the preset threshold.

[0053] The expected dose value is obtained by time integration of a nonlinear function that deviates the predicted mean sequence from the reference value; The expected dose value is the expected future dose or the expected total dose. The expected future dose is the value obtained by integrating the nonlinear function of the predicted mean sequence deviating from the reference value over time; the expected total dose is the sum of the expected future dose and the cumulative equivalent causal dose.

[0054] The pass test condition corresponding to the total dose expectation is that the total dose expectation is higher than the dose limit.

[0055] The pass test condition corresponding to the expected future dose is the difference between the expected future dose and the upper limit of the dose and the cumulative equivalent efficacy dose.

[0056] The overtest protection actions include any one or more of the following: reducing the rate of temperature rise or fall, reducing vibration intensity, shortening the holding time, and entering a conservative operating condition phase or terminating the test and indicating that the dose limit has been reached.

[0057] An environmental test abnormality interruption early warning and over-test protection device, applied to an environmental test abnormality interruption early warning and over-test protection method, includes: The data acquisition module is used to acquire multi-source time-series data of the environmental test process and control command data that is time-synchronized with the multi-source time-series data; The preprocessing and construction module is used to preprocess the multi-source time series data and construct the model input sequence by combining the control command data from multiple past time points with the preprocessed multi-source time series data. The probabilistic time series prediction module is used to input the model input sequence into the probabilistic time series prediction model to predict multi-step prediction results for multi-source time series data. The risk indicator calculation module is used to calculate risk indicators based on the actual observation data at the current moment of multi-source time series data and the results of multi-step future predictions. The exception module is used for exception detection and handling. The pass test evaluation module is used to perform pass test evaluation operations.

[0058] An environmental test abnormal interruption early warning and overtest protection device also includes a time synchronization module and an adaptive threshold calculation module.

[0059] The risk indicator calculation module includes an anomaly scoring module and an over-limit probability calculation and risk fusion module.

[0060] The anomaly module includes a drift separation and anomaly source determination module, a control and protection execution module, and an event logging and reporting module.

[0061] An environmental test abnormal interruption early warning and overtest protection device interacts with a data acquisition module via a message bus or shared storage; the data transmitted or stored in the message bus or shared storage adopts a data frame structure containing timestamps, channel identifiers, numerical values, units, and quality markers; the control and protection execution module provides a safety interlock interface with the test equipment controller for performing at least one of the following steps: power failure, shutdown, ventilation / inertization, and locking.

[0062] An environmental test abnormality interruption early warning and overtest protection system includes: environmental test equipment, sensor group, test equipment controller, and environmental test abnormality interruption early warning and overtest protection device; the environmental test abnormality interruption early warning and overtest protection device is communicatively connected or electrically interlocked with the test equipment controller.

[0063] This allows for automatic execution of protective measures when a genuine anomaly is detected, and outputs maintenance recommendations when sensor drift is detected.

[0064] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of an environmental test abnormal interruption warning and overtest protection method.

[0065] It utilizes multi-source time-series data for probability prediction and trend early warning; it can classify sensor drift and real anomalies; it also enables overtest management; and improves the safety and effectiveness of experiments.

[0066] Example 2: like Figure 1 As shown, Embodiment 2 has all the features of Embodiment 1, except that: System composition, data interface and time synchronization: The environmental testing system includes: environmental testing equipment (temperature and humidity chamber / vibration table, etc.), sensor group, test equipment controller, and computing and control device (which can be an industrial PC or edge computing module).

[0067] Discrete sampling time (in The sampling period is denoted as (using a non-negative integer time index) and the unit is denoted as . (in (The time interval between adjacent sampling moments) forms a frame of synchronization data: , in, For a moment Data frames; For timestamps; This refers to the multi-source observation vector corresponding to the multi-source time series data (which includes at least environmental and equipment status variables). For the control command / actuator state vector corresponding to the control command data (such as set value, heating duty cycle, fan speed, etc.); This is a quality marker vector (used to indicate missing, over-range, calibration status, communication anomalies, etc.).

[0068] Environmental test abnormality interruption early warning and over-test protection devices read data via message bus or shared storage. And perform time alignment: If multiple sampling rate channels exist, resample to the same reference clock. Priority is given to ensuring the time synchronization accuracy of safety-related channels (temperature, power, smoke, etc.).

[0069] Preprocessing and sliding window construction: Let the multi-source observation vector be... ,in To observe the number of channels, For a moment The Each channel observation value For channel indexing, This represents the transpose operation. Perform at least one of the following preprocessing steps: missing data repair, outlier suppression, robust normalization, differential feature construction, and... Time alignment.

[0070] The extended feature vector is obtained by concatenating the multi-source observation vector with the control command / actuator state vector. , in To extend the feature vector, ";" indicates vector concatenation.

[0071] The multi-source time-series data includes at least one multimodal data source, such as images, audio, or thermal images. The probabilistic time-series prediction model extracts features from the multimodal data and fuses them with an extended feature vector for prediction and early warning.

[0072] Multimodal extension: Introducing multimodal data such as environmentally resistant cameras / thermal imaging / audio, extracting features, and then combining them with... The system integrates and maintains the closed-loop framework of "probabilistic prediction - residual and out-of-limit probability - drift separation - threshold adaptation - protection and reporting".

[0073] Let the window length be (in Construct the window input sequence (for the number of sampling points contained in the sliding window): , in For a moment The window input sequence.

[0074] By analogy, we can obtain , The input sequence for the model is derived by using extended feature vectors from multiple past time points.

[0075] Probabilistic time series prediction models and training: The probabilistic time series prediction model is as follows: , In the formula, Input sequence to the model, This represents a probabilistic time series forecasting model. For model parameters, To address the present Multi-source time series data The predicted mean vector; To address the present Multi-source time series data The prediction uncertainty parameter; To predict the step size; Indicates from Time's up The predicted mean sequence of multi-source time series data at any given time. Indicates from Time's up The sequence of prediction uncertainty parameters at time points. It is a positive integer.

[0076] In this embodiment, the prediction uncertainty parameter is the covariance matrix.

[0077] Probabilistic time series forecasting models at time 1 Based on the previous window Output the predicted distribution parameters for the current observation and multiple future steps.

[0078] The probabilistic time series forecasting model structure can adopt LSTM / GRU / Transformer. Preferably, an attention-based Transformer is used to enhance the modeling ability for long-term dependencies and operating condition changes; the prediction uncertainty is in the form of heteroscedasticity, so that the model outputs greater uncertainty when the disturbance increases or the operating condition changes, thereby reducing misjudgment.

[0079] For offline training, negative log-likelihood loss is used.

[0080] The formula for the negative log-likelihood loss is: , in, For training the loss function; A set of time indices for training samples; For variable indexing; To predict the mean vector The One component; For the first The predictive standard deviation of each component; It is the natural logarithm function.

[0081] The predicted standard deviation is from The square root of the diagonal elements is obtained or directly output by a probabilistic time series prediction model.

[0082] Standardized residuals, outlier scoring, and trend enhancement: When multi-source time series data are the actual observation data at the current moment Upon arrival, the standardized residuals are calculated.

[0083] The formula for calculating the standardized residual is: , In the formula, For standardized residual vectors; Indicates to Inverse square root matrix operations, This refers to the actual observation data at the current moment in the multi-source time series data. To address the present Multi-source time series data The predicted mean vector.

[0084] When predicting uncertainty parameters using a diagonal approximation of the covariance matrix, The equivalent is element-wise normalization based on the channel standard deviation.

[0085] Anomaly scores are constructed based on Mahalanobis distance. The formula for calculating the anomaly score is as follows: , In the formula, This is for anomaly scoring.

[0086] To amplify trend anomalies rather than transient noise, anomaly scores are subjected to exponentially weighted smoothing. The formula for calculating exponentially weighted smoothing is: , In the formula, for Outlier scores after time-smoothing; for Outlier scores after time-smoothing The smoothing coefficient satisfies , The anomaly score before smoothing.

[0087] Integration of multi-step over-limit probability and risk indicators: For each controlled variable Set security limits (in For variables (safety ceiling), calculated based on predicted distribution for the future Probability of exceeding the limit within a step: , in, For variables In the future The probability of exceeding the limit within a step; For prediction step index ( ); To predict the mean vector The One component; This corresponds to the standard deviation of the forecast; The cumulative distribution function of the standard normal distribution; This is the chain multiplication operator.

[0088] Sum the probabilities of each variable exceeding the limit according to their weights: , in, To summarize the probability of exceeding the limit, that is, the probability of exceeding the limit in multiple future steps; For variables The weighting coefficients satisfy Furthermore, it can be normalized.

[0089] The formula for calculating the risk indicator is as follows: , In the formula, As a risk indicator; To integrate the weighting coefficients, satisfy the following conditions: ; For abnormal rating scale parameters; It is an exponential function. Preferred. .

[0090] The above design ensures that: as the standardized residual continues to increase ( (Increase) or the probability of exceeding the limit in the future ( When the risk indicator increases, This means that the warning is triggered in advance, before the limit is exceeded.

[0091] Drift separation and anomaly source classification: For each channel Construct exponential smoothing with both fast and slow time scales: , In the formula, for Time Channel The fast time-scale smoothed value; for Time Channel Slow time-scale smoothed values; The smoothing coefficient is and satisfies , for Time Channel Fast time-scale smoothing value, for -1 Time Channel Slow time-scale smoothing value, For the present Multi-source time series data channels The data.

[0092] The drift index is: , In the formula, For channel The drift index; This represents the absolute value operation; To prevent zero constant; Channel within the current window The median absolute deviation is used as the robustness of dispersion.

[0093] To avoid misclassifying single-channel drift as a genuine anomaly, a cross-sensor correlation consistency metric is introduced. The correlation structure matrix for the current window is calculated. and correlation matrix with normal baseline Compare: , In the formula, As a measure of relevant consistency differences; It is the Frobenius norm; It is obtained from offline statistics of historical normal data and can be grouped according to working conditions.

[0094] A preferred criterion for anomaly source classification is: let the drift threshold be... (in (For drift determination threshold), the relevant difference threshold is: (in (For the relevant consistency difference threshold), when a certain channel exists satisfy and When, it is determined to be sensor drift; when If the residuals of multiple channels increase simultaneously, it is determined that the test process is abnormal or the test object is abnormal.

[0095] The synchronous increase of multi-channel residuals can be caused by The determination is achieved when multiple components continuously deviate from zero.

[0096] Adaptive threshold setting and online updates: Because the statistical characteristics differ across different test phases (heating, isothermal, cooling, cycling, etc.), quantile thresholds grouped by operating condition are used. The operating condition label is denoted as... (in For the operating condition category determined by setpoints, control states, or test procedures, thresholds are set using the conditional quantiles of risk indicators from historical normal data: , In the formula, An adaptive threshold; Represents the quantile operator; The quantile level (e.g., 0.99 or 0.995); This represents a random variable representing a risk indicator within normal data.

[0097] Online updates can be performed for each operating condition. Maintain risk indicators The sliding sample set is periodically updated. Alternatively, extreme value theory can be used to fit and update the tail.

[0098] Tail threshold enhancement: The tail distribution of the adaptive threshold is fitted and updated using extreme value theory to stably control an extremely low false alarm rate.

[0099] Overtest dose assessment and protection: The equivalent stress "instantaneous dose rate" is defined, and the formula for calculating the equivalent stress dose rate is as follows: , In the formula, For a moment The equivalent effect dose rate; For variables Dose weighting; For variables Reference values ​​(such as normal temperature and humidity or target set values); For variables Scale parameters; It is a non-linear exponent and satisfies , For the present Multi-source time series data variables The data.

[0100] The formula for calculating cumulative equivalent stress dose is: , In the formula, for Accumulated equivalent force dose over time; for Accumulated equivalent force dose over time; The sampling period is For a moment The equivalent force dose rate.

[0101] Let the upper limit of the dose be (in (For the maximum permissible cumulative dose), the remaining safety margin is defined as: , In the formula, This represents the remaining safety margin.

[0102] To achieve prospective overtest protection, the expected future dose is estimated using the predictive distribution.

[0103] The formula for calculating the expected future dose is: , The formula for calculating the expected total dose is: , In the formula, Indicates the expected total dose; The total expected dose, which is the future Cumulative dose expectation after step express Expected dose at any given time.

[0104] In engineering implementation, The predicted mean can be used as an approximation. The formula for calculating the expected dose at time t is: , In the formula To predict the mean vector The Each component.

[0105] when Less than the preset threshold or During testing, protective measures were taken, including reducing stress intensity, reducing the temperature / vibration slope, shortening the holding time, switching to a conservative phase, or terminating the test and indicating that the upper dose limit had been reached.

[0106] Decision-making logic, protective actions, and event tracing: when Furthermore, when the source of the anomaly is determined to be an abnormality in the test process or an abnormality in the test object, an anomaly interruption warning is immediately triggered and protective actions are executed according to priority, such as: cutting off the heating circuit, stopping the vibration output, disconnecting the main power contactor, starting the exhaust / inerting process, locking the test steps to a safe state, and sending a notification to the host computer / remote platform.

[0107] When the source of the anomaly is determined to be sensor drift, a drift alarm is output and maintenance suggestions are generated: calibrate / replace the sensor, check the wiring and installation location; at the same time, the drift channel can be downweighted or a redundant sensor channel can be switched in the risk fusion (if it exists).

[0108] To ensure traceability, data snapshots are saved before and after the anomaly. Let the snapshot half-window length be... (in (Snapshot half-window length in terms of sampling points), storage interval within as well as The system automatically generates event reports based on the classification results of anomaly sources and the execution status of protection actions.

[0109] An environmental test abnormality interruption early warning and overtest protection method and device can accomplish the following: Based on the trend-based early warning of predicted residuals and probability of exceeding limits, an early warning is issued and interlocked protection is implemented before the limit is exceeded. Reliably distinguish sensor drift from anomalies in the actual test process / anomalies of the tested object, thereby reducing false alarms and missed alarms; The test equipment and the test object are protected in a coordinated manner, and the risk of overtesting is managed in a dose-based manner and intervened in advance. Generate event snapshots and reports to support traceability and maintenance decisions.

[0110] An environmental testing abnormality interruption early warning and overtest protection method and device can: Early warnings can be issued based on trend residuals and the probability of exceeding limits, significantly improving the lead time and safety of handling. Drift separation reduces false alarms, missed alarms, and unnecessary interruptions caused by sensor drift, thereby improving the effective output of experiments; Risk indicator integration unifies multi-channel and multi-condition information into an interpretable risk quantity, which facilitates interlocking control. Reduce the risk of accidental damage to expensive prototypes due to overtesting through dosage and margin management; Event snapshots and reports improve the efficiency of tracing and maintenance; automatically retain data snapshots before and after anomalies to generate reports.

[0111] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for early warning of abnormal interruption and protection against overtesting in environmental testing, characterized in that: Includes the following steps: S1: Acquire multi-source time-series data of the environmental test process and control command data that is time-synchronized with the multi-source time-series data; The multi-source time-series data includes environmental sensor data and experimental equipment operating status data; S2: Preprocess the multi-source time series data, and construct the model input sequence by combining the control command data from multiple past time points with the preprocessed multi-source time series data; Step S2 constructs the model input sequence using a sliding window method; S3: Input the input sequence of the model into the probabilistic time series prediction model to predict the multi-source time series data and obtain the prediction results for the future multiple steps; The future multi-step prediction results include a predicted mean sequence and a prediction uncertainty parameter; The prediction uncertainty parameter is one or more of variance, covariance matrix, and diagonal approximation of covariance matrix; The probabilistic time-series prediction model is any one of Long Short-Term Memory Network, Gated Recurrent Unit Network, or Transformer Network based on attention mechanism; The output probability distribution of the probabilistic time series prediction model is a Gaussian distribution or a Student's t-distribution. The probabilistic time series prediction model is trained by minimizing the negative log-likelihood loss; S4: Risk indicators are obtained by calculating risks based on the actual observation data at the current moment and the results of multi-step predictions in the future from multi-source time series data. The risk indicator is obtained by fusing anomaly scores with the probability of future multi-step exceedances. The risk indicator is positively correlated with the anomaly score, and the risk indicator is positively correlated with the probability of exceeding the limit in multiple future steps; The anomaly score is calculated based on the actual observation data at the current moment of multi-source time series data and the results of multi-step prediction in the future. The method for calculating the anomaly score is to first calculate the standardized residuals and then construct the anomaly score based on the standardized residuals. The anomaly score is constructed by aggregating multi-channel residuals in Mahalanobis distance form based on the standardized residuals and performing trend enhancement operations. The future multi-step exceedance probability is obtained by weighted summation of the future exceedance probabilities of each variable calculated based on the future multi-step prediction results. S5: If the risk indicator exceeds the risk threshold, an anomaly judgment and handling operation will be performed; otherwise, an overtest evaluation operation will be performed. The anomaly judgment and handling operation is to perform an anomaly source judgment operation; when the anomaly source is a real anomaly, an anomaly interruption warning is output and a protection action is executed; when the anomaly source is sensor drift, a drift alarm is output.

2. The method for early warning of abnormal interruption and overtest protection in environmental testing according to claim 1, characterized in that: The risk threshold is an adaptive threshold; the adaptive threshold is set based on historical normal data and combined with the current working conditions to set the risk indicators.

3. The method for early warning of abnormal interruption and overtest protection in environmental testing according to claim 2, characterized in that: The adaptive threshold is obtained by estimating the quantiles of the risk indicators in historical normal data by grouping them according to operating conditions, and is updated online as the control command data or test phase changes.

4. The method for early warning of abnormal interruption and overtest protection in environmental testing according to claim 1, characterized in that: The protective actions include any one or more of the following: cutting off the power supply to the heating circuit, turning off the main power relay, stopping the vibration output, starting the exhaust, starting the inerting process, locking the test steps and maintaining a safe state, and triggering an audible and visual alarm and remote notification. The anomaly source determination operation constructs a sensor drift index based on multi-source time-series data using a dual-time-scale statistical approach, and obtains a cross-sensor correlation consistency measure through correlation calculation. Then, the anomaly sources are classified according to the sensor drift index and the correlation consistency measure. The sensor drift index is constructed from the difference between the fast timescale smoothed value and the slow timescale smoothed value, and normalized by robust dispersion; the cross-sensor correlation consistency measure is obtained by comparing the correlation structure of multiple past time points with the normal baseline correlation structure.

5. The method for early warning of abnormal interruption and overtest protection in environmental testing according to claim 1, characterized in that: The overtest assessment operation is to perform overtest protection actions when the assessment indicators meet the overtest conditions; the assessment indicators are one or both of the remaining safety margin and the expected dose value. The remaining safety margin is calculated by subtracting the cumulative equivalent effect dose from the upper dose limit; The cumulative equivalent stress dose is the value obtained by integrating the nonlinear function of the multi-source time-series data deviating from the reference value over time. The expected dose value is obtained by time integration of a nonlinear function that deviates the predicted mean sequence from the reference value; The overtest protection actions include any one or more of the following: reducing the heating rate, reducing the cooling rate, reducing vibration intensity, shortening the holding time, entering a conservative operating condition phase, stopping the test and indicating that the upper dose limit has been reached.

6. An environmental test abnormality interruption early warning and overtest protection device, applied to the environmental test abnormality interruption early warning and overtest protection method according to any one of claims 1 to 5, characterized in that: include: The data acquisition module is used to acquire multi-source time-series data of the environmental test process and control command data that is time-synchronized with the multi-source time-series data; The preprocessing and construction module is used to preprocess the multi-source time series data and construct the model input sequence by combining the control command data from multiple past time points with the preprocessed multi-source time series data. The probabilistic time series prediction module is used to input the model input sequence into the probabilistic time series prediction model to predict multi-step prediction results for multi-source time series data. The risk indicator calculation module is used to calculate risk indicators based on the actual observation data at the current moment of multi-source time series data and the results of multi-step future predictions. The exception module is used for exception detection and handling. The pass test evaluation module is used to perform pass test evaluation operations.

7. An environmental testing abnormality interruption early warning and overtest protection system, characterized in that: include: The environmental testing equipment, sensor group, testing equipment controller, and the environmental testing abnormal interruption early warning and overtest protection device as described in claim 6; the environmental testing abnormal interruption early warning and overtest protection device is communicatively connected or electrically interlocked with the testing equipment controller.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the environmental test abnormal interruption warning and overtest protection method as described in any one of claims 1 to 5.

Images