Calibration full life cycle prediction system and method for crash test dummy

Abstract

The invention belongs to the technical field of dummy detection, and particularly relates to a collision test dummy-oriented calibration full life cycle prediction system, which comprises a data source layer, a data layer, a core layer and an application layer, the data source layer covers dummy calibration full-scene data; the data layer accesses multi-source data through an acquisition and storage module and preprocesses and stores the multi-source data, and a time sequence construction module constructs a time sequence data set according to dummy IDs and component IDs; the core layer completes performance degradation modeling and residual available life prediction based on a multi-model fusion algorithm; and the application layer receives the prediction result, automatically updates a dummy state, pushes a spare part purchasing demand and visually displays the spare part purchasing demand. The problems that dummy calibration management depends on manual operation, and systematic data integration and intelligent analysis means are lacked in the prior art can be solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of dummy detection, and particularly relates to a calibration full life cycle prediction system and method for a crash test dummy. BACKGROUND

[0002] In the field of vehicle safety performance evaluation, crash test is one of the core verification methods, and the dummy is a key test tool for simulating the mechanical response and physiological signal change of the human body in the crash process. The accuracy of the sensor data of the dummy directly determines the reliability and credibility of the evaluation results of the vehicle safety performance. According to international and industry related standards such as ISO, FMVSS, etc., the dummy needs to complete the calibration operation according to the specified period, and through the performance calibration of the built-in accelerometer, force sensor, displacement sensor and other core components, it is ensured that the deviation of the sensor output data from the preset standard value is within the allowed range, thereby ensuring the effectiveness of the data of each crash test.

[0003] However, the current management mode of the dummy calibration in the industry still mainly relies on manual recording and fixed period inspection. This mode has significant defects in actual application and is difficult to meet the needs of modern automobile crash test for efficiency, accuracy and traceability. The specific problems are as follows: 1. Manual management relies on the memory and record of the number of times the dummy is used and the last calibration time by the staff. Due to human negligence, information transmission deviation and other factors, the dummy may be used beyond the specified calibration period; 2. The service life of the dummy is closely related to the recession state of the core components. However, under the existing management mode, the calibration records are scattered in paper documents or independent electronic spreadsheets, and lack of systematic integration; 3. The performance recession of the core components of the dummy has periodic and progressive characteristics. The recession degree is strongly related to the use frequency, impact energy and environmental conditions. However, the existing technology cannot quantitatively model the relationship between these influencing factors and the recession degree; 4. The existing technology does not use AI to analyze the recession trend of the calibration data, and cannot provide early warning.

[0004] In summary, the current dummy calibration management mode relies on manual operation, lacks systematic data integration and intelligent analysis means, and urgently needs a technical solution that can realize the full life cycle tracking, recession trend quantitative prediction and spare parts collaborative management of the dummy calibration. SUMMARY

[0005] The technical problem solved by the present application is to provide a calibration full life cycle prediction system and method for a crash test dummy, to solve the problem of manual operation, lack of systematic data integration and intelligent analysis means in the existing dummy calibration management.

[0006] The basic solution provided by this invention is a calibration lifecycle prediction system for crash test dummies, comprising a data source layer, a data layer, a core layer, and an application layer, wherein: The data source layer serves as the data input source, covering data from the entire scenario of dummy calibration. The data layer includes a data acquisition and storage module and a time series construction module. The data acquisition and storage module is used to access the data source layer, collect multi-source data from dummies, and store it after preprocessing. The time series construction module is used to construct a time series dataset from the preprocessed multi-source data of dummies according to the dummy ID and component ID. The core layer receives time-series datasets, models the performance degradation of dummies and components, and predicts their remaining usable life based on a multi-model fusion algorithm, then outputs the prediction results. The application layer is used to receive the prediction results output by the core layer, automatically update the dummy status, push spare parts procurement requirements, and display them visually. Specifically, the time series construction module constructs a time series dataset from the preprocessed multi-source dummy data according to dummy ID and component ID as follows: Based on the business scenario of dummy calibration and the input requirements of multi-model fusion algorithm, basic identification features, working condition quantification features, decay trend features and environmental correlation features are extracted from preprocessed multi-source dummy data. Based on benchmark time axis anchoring and data interpolation completion, time-series alignment and feature data fusion processing are performed on multi-source dummy data to generate fused time-series data. Based on the input characteristics of each model in the multi-model fusion algorithm, the fused time series data is optimized for differences, generating multiple sets of model-specific datasets; Version metadata is embedded in each model's unique dataset, and a mapping relationship is established with the model version. This metadata is then stored in a database containing a 3D index that includes dummy ID, component ID, and model type.

[0007] Furthermore, the process of performing time-series alignment and feature data fusion on the multi-source dummy data based on reference time axis anchoring and data interpolation to generate fused time-series data specifically involves: Using the time of the calibrated event as the core anchor point, a baseline timeline is constructed according to a preset minimum time unit; The collected dummy multi-source data is divided into high-frequency sensor data, low-frequency data, and event-triggered data. The high-frequency sensor data is aggregated according to the calibration event time window and associated with the low-frequency data of the calibration event. The event-triggered data is anchored to the reference time axis according to the occurrence time axis and marked as event labels. The basic identification features, working condition quantification features, decay trend features, and environmental correlation features of the same dummy component are concatenated in chronological order to form a high-dimensional time-series vector, which represents the fused time-series data.

[0008] Furthermore, the core layer includes a model library module, a training module, an inference module, a fusion module, an uncertainty module, and an interpretation module, wherein: The model library module is used to integrate statistical models, machine learning models, and deep learning models for predicting the lifespan of dummy parts, and to automatically select models based on the data characteristics of time series datasets and to manually specify the models to select. The training module is used to extract data features from the time series dataset called from the data layer, load the target model in the model library module, iterate the training using the training set of historical data according to the preset optimizer and loss function, calculate the model evaluation index using the validation set of historical data, and generate a standardized model package. The inference module receives the time series dataset from the data layer, calls the corresponding standardized model package trained by the training module to perform inference on the time series dataset, and outputs the RUL prediction results, failure probability curves and model version identifier. The fusion module selects multiple optimal models based on the evaluation accuracy in the training module, and then uses the latest validation set of the optimal models. After calculating the weights of each model, RUL prediction and prediction results are fused to obtain the fused model package. The uncertainty module is used to resample the input time series dataset with replacement to generate a new feature dataset, and to perform RUL prediction on the new feature dataset according to the fused model package. The confidence intervals of the output RUL prediction values ​​are extracted and the confidence intervals are verified. The inference module is also used to adjust the called model based on the verification results; The explanation module is used to calculate the influence of each feature on the RUL prediction result based on the RUL prediction result output by the inference module and the corresponding input features, and to select several key features and generate an adjustment scheme for each key feature.

[0009] Furthermore, the uncertainty module resamples the input time series dataset with replacement to generate a new feature dataset, and performs RUL prediction on the new feature dataset according to the fused model package. Confidence intervals are extracted from the output RUL prediction values, and the confidence intervals are validated as follows: The time series features in the input time series dataset are resampled with replacement to generate a new feature dataset. The fused model package is called to perform RUL prediction on the new feature dataset, and new RUL prediction results are obtained. Sort the new RUL prediction results from smallest to largest, take the 2.5th percentile as the lower limit of the confidence interval, and the 97.5th percentile as the upper limit of the confidence interval; The confidence interval is validated for its width, and the validation result is fed back to the inference module to help the inference model optimize its calling strategy.

[0010] Furthermore, the explanation module calculates the influence of each feature on the RUL prediction result based on the RUL prediction result output by the inference module and the corresponding input features, and selects several key features. For each key feature, an adjustment scheme is generated as follows: Obtain the model version identifier, RUL prediction results, and input feature dataset transmitted by the inference module; The model type is identified by the model version identifier, and different preset feature influence value calculation methods are called to calculate the corresponding feature influence values ​​according to different model types. The characteristic influence values ​​calculated by different methods are normalized to The intervals are divided into levels according to preset thresholds, associated with business names, and a feature importance ranking table is output. Based on the feature importance ranking table, the top N key impact features are extracted. Pre-defined business constraint rules are then applied to each key impact feature according to a preset adjustment range to generate candidate adjustment values. Simultaneously, values ​​violating the business constraint rules are removed, and the remaining candidate adjustment values ​​are retained. The expression is as follows:

[0011] in, Candidate adjustment values, , , To define the range of feature values ​​allowed by the preset business constraint rules, Let be the range of values ​​for the i-th feature. Preset adjustment range; Keeping the feature values ​​of the remaining features unchanged, we replace the feature values ​​of the key features with candidate adjustment values ​​to construct multiple candidate feature vectors. The inference module is invoked, the candidate feature vector is input into the model package of the original inference module, the corresponding RUL prediction result is output, and the RUL change is calculated. Based on a preset priority scoring function, the priority score is calculated according to the change in RUL, expressed as follows:

[0012] in, The change in RUL Rate the priority. To adjust the implementation cost of the solution, it is set according to the business scenario; Priority rating Sort in descending order and filter. The solution is to retain the top N priority schemes.

[0013] Furthermore, the model type is identified based on the model version identifier, and different preset feature influence value calculation methods are called to calculate the corresponding feature influence values ​​according to different model types. Specifically: If the model is a statistical model, the characteristic influence value is calculated based on the parameter influence coefficient method, and the expression is:

[0014] in, For the first The model coefficients for each feature, For the first The range of values ​​for each feature; For the statistical model corresponding to the first The influence value of each feature; If the model is a machine learning model, the feature influence value is calculated based on the feature gain method, and the expression is:

[0015] in, This represents the influence value of the i-th feature in the machine learning model. The total information gain of the i-th feature is... The sum of the total information gains of the n features; If the model is a deep learning model, the feature influence value is calculated based on the SHAP value method, and the expression is:

[0016] in, For SHAP value mapping functions, Let be the range of values ​​for the i-th feature; This represents the influence value of the i-th feature in the deep learning model.

[0017] Furthermore, the application layer includes a dummy and component status automatic update module, a spare parts requirement generation module, and a visualization display module, wherein: The automatic status update module for dummies and components is used to construct a four-level status system of normal, early warning, emergency, and failure based on the prediction results output by the core layer, and to automatically update the status of dummies and components. The spare parts requirement generation module is used to build a procurement requirement calculation model based on the prediction results of the core layer and spare parts data, and output a standardized procurement requirement sheet. The visualization module is used to visualize the core layer prediction process, prediction results, the status of dummies and components, and standardized procurement requirements.

[0018] A calibration lifecycle prediction method for crash test dummies, applied to the aforementioned calibration lifecycle prediction system for crash test dummies, includes: S1: Receive full-scene data for dummy calibration, preprocess and store it to generate preprocessed multi-source dummy data; S2: Construct a time series dataset from the preprocessed multi-source dummy data according to dummy ID and component ID; S3: Input the time series dataset into the multi-model fusion algorithm to model the performance degradation of the dummy and its components and predict the remaining usable life, and output the prediction results; S4: Based on the output prediction results, automatically update the dummy status, push spare parts procurement requirements, and display them visually.

[0019] The principle and advantages of this invention are as follows: The technical solution of this application realizes the full life cycle management of crash test dummy calibration through "layered architecture + multi-model fusion"; wherein, the data source layer covers the full scenario data of dummy calibration, the data layer first preprocesses the multi-source data and constructs a time series dataset according to dummy ID and component ID, and extracts four types of features such as basic identifiers and working condition quantification, constructs a benchmark time series axis with the calibration event time as the anchor point, aligns high-frequency sensor, low-frequency and event-triggered data and fuses them into a high-dimensional time series vector, and also optimizes and generates a dedicated dataset for multi-model input characteristics and establishes a three-dimensional index storage; the core layer integrates statistical, machine learning and deep learning models, generates a standardized model package through iterative training by the training module, outputs results such as remaining usable life (RUL) by the inference module, fuses the optimal model prediction results according to the model accuracy weight, the uncertainty module verifies the confidence interval through resampling, and the interpretation module calculates the feature influence degree and generates adjustment schemes; the application layer updates the dummy state with a four-level state system based on the prediction results of the core layer, constructs a procurement model to generate a demand order and displays it visually.

[0020] The advantages are: it effectively solves the problems of traditional dummy calibration and management relying on manual labor and lacking system integration and intelligent analysis. On the one hand, by constructing time series datasets and multi-model fusion algorithms, accurate modeling and RUL prediction of the performance degradation of dummies and components can be achieved. Combined with uncertainty verification, the prediction credibility can be improved. The interpretation module can also locate key influencing features and provide adjustment schemes to avoid the dummies from being used beyond their service life. On the other hand, the application layer automatically updates the dummy status and pushes spare parts procurement requirements, replacing manual recording and decentralized management, improving the efficiency and traceability of dummy calibration management, ensuring the validity of vehicle crash test data, and meeting the requirements of modern vehicle safety performance assessment for accuracy and intelligence. Attached Figure Description

[0021] Figure 1This is a system architecture diagram of an embodiment of the present invention; Figure 2 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0022] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1 As shown: A calibration lifecycle prediction system for crash test dummies includes a data source layer, a data layer, a core layer, and an application layer, wherein: The data source layer serves as the data input source, covering data from the entire dummy calibration scenario. In this embodiment, the data source layer primarily uses automated data acquisition, supplemented by manual data entry. Specifically: 1. Hardware devices such as the dummy calibration platform, sensor data acquisition instrument, and environmental monitoring equipment are directly connected to the data source layer via industrial Ethernet or RS485 bus, configuring data acquisition interfaces to achieve real-time automatic acquisition of raw sensor data and calibration condition data; 2. It interfaces with the automotive crash test management system via API to automatically synchronize usage record data for each test, and simultaneously interfaces with the enterprise ERP system to automatically obtain real-time updates of spare parts inventory data and historical procurement data; 3. Low-power IoT tags are installed on the dummy and core components to record data such as the dummy's location, movement trajectory, and idle time, supplementing static state data not covered by the test system during use; 4. For other static data that cannot be collected, such as the dummy's manufacturing date, component calibration certificate number, etc., and non-routine operation records, the operator enters them at the system front end.

[0023] Therefore, when the above data source layer accesses various data sources, the total data covered for the entire scenario of dummy calibration includes basic identity information of the dummy and components, dynamic data of the calibration process, data related to degradation during use, reference data of standards and thresholds, and data related to spare parts and procurement; thus achieving multi-scenario and multi-dimensional data coverage.

[0024] The data layer includes a data acquisition and storage module and a time series construction module. The data acquisition and storage module is used to access the data source layer, collect multi-source data from dummies, and store it after preprocessing. The time series construction module is used to construct a time series dataset from the preprocessed multi-source data of dummies according to the dummy ID and component ID. The data acquisition module employs protocol adaptation and interface standardization technologies to achieve seamless access to all types of data in the data source layer. The acquired multi-source dummy data first undergoes data cleaning, which includes noise filtering, missing data repair, and data consistency verification. Noise filtering targets random noise in high-frequency sensor data (such as accelerometer data) using a combined algorithm of "sliding window smoothing + threshold removal." First, a moving average smoothing process is performed using 10 sampling points as a window to reduce high-frequency noise interference; then, based on... The criterion (3 standard deviations under normal distribution) sets a noise threshold and calculates the data mean. with standard deviation It will exceed Outliers in the interval are marked as noise and replaced with the linear interpolation result of two adjacent valid data points to ensure the smoothness of high-frequency data (smoothing reduces data fluctuation). For falsely triggered data in event-triggered data (such as calibration anomaly alarm signals), "event correlation verification" technology is used to compare the calibration equipment status data at the time the alarm signal occurred (such as whether the calibration table loading force exceeded the normal range). If the equipment status is normal, it is determined to be a false trigger, and the alarm data is automatically removed. The false trigger data removal rate is [percentage missing]. ; Missing data repair employs a differentiated repair strategy based on differences in missing rates. When the missing rate of a single data entry... (For example, if only humidity data is missing in a calibration record), the "neighborhood interpolation method" is used for repair. This means that high-frequency sensor data is filled with linear interpolation of adjacent timestamp data, while low-frequency operating condition data (such as temperature and humidity) is filled with the mean of the calibration data from the same batch. When the missing rate is within a certain range... If there is a gap (e.g., 8 minutes of displacement data are missing within a 10-minute period), a "filling method based on historical similar working conditions" is used to search the historical database for similarity with the current calibration working conditions (e.g., temperature, loading speed, dummy model). The records are used to extract displacement data for the corresponding time period as a reference, and the data is filled by weighted averaging (the higher the similarity, the greater the weight); when the missing rate If the data is marked as "incomplete," an alarm mechanism is triggered, and the person responsible for data collection (such as the equipment administrator or operator) is notified via system message. The cause of the missing data is verified, and supplementary data is collected. The data integrity is then restored after the missing data is repaired. .

[0025] Data consistency verification employs a "cross-source comparison + logical verification" technique to eliminate data inconsistencies. Regarding cross-source comparison, the comparison... The location data of the dummy collected by the terminal and the location of the dummy used recorded by the test system (e.g., The dummy is displayed in the "storage area," but the test system records it as "participating in the test." If there is an inconsistency, the dummy's movement trajectory data will be retrieved. The location time series of the terminal and the start / end timestamps of the test were used to determine the cause of the data discrepancy (e.g., terminal location time series) and the test start / end timestamps. (Label upload delay), data is corrected based on the data source with the most recent time; in terms of logical verification, verification logic is built based on business rules (such as "the end time of calibration records must be later than the start time" and "the sensor accuracy level must match the component model, such as..."). The accuracy class of the dummy head accelerometer needs to be [missing information]. "),pass The query statement is used to batch detect contradictory data, automatically correct logical errors (such as swapping end and start times), and mark those that cannot be automatically corrected as "awaiting manual review". Data consistency verification coverage is also checked. .

[0026] After data cleaning, data standardization is performed, including format standardization and numerical standardization. Format standardization uses a "field mapping + unit unification" technique to convert heterogeneous data from different data sources into a unified format. Numerical standardization addresses the magnitude differences of different types of feature data by using the Z-Score normalization algorithm to normalize numerical data, eliminating the impact of magnitude differences on subsequent model training. Numerical standardization coverage... .

[0027] After data standardization, the multi-source dummy data employs a "hybrid storage architecture + three-dimensional indexing" technology to achieve efficient storage and fast retrieval. The hybrid storage architecture includes high-frequency real-time data storage, structured business data storage, unstructured data storage, and cold data archiving. High-frequency real-time data storage utilizes the InfluxDB time-series database to store high-frequency sensor data. This database is optimized for time-series data, supporting high write throughput and time-range queries. Data is managed through a "time partitioning + data retention strategy," which involves partitioning data according to calibrated event times (e.g., by day / month) and setting data retention periods (e.g., retaining high-frequency raw data for 2 years, automatically archiving to cold storage after the period expires) to ensure timely real-time data query responses. .

[0028] Structured business data storage utilizes a MySQL relational database to store basic identity data for dummies and components (dummy ID, model, component ID, accuracy level), calibration condition data (temperature, humidity, operator ID), spare parts inventory data, and other structured data. A standardized database table structure is designed, and data consistency is achieved through foreign key relationships (e.g., the "dummy_id" field in the "calibration record table" is linked to the "dummy_id" primary key in the "dummy basic information table"). Complex SQL queries are supported, and query response time is [not specified]. .

[0029] Unstructured data storage utilizes the MinIO object storage system to store calibration report PDFs, component damage photos, maintenance videos, and other unstructured data. A unique "data identifier ID" (formatted as "data type-dummy ID-timestamp," e.g., "report-THOR2023001-20240510") is generated for each unstructured data item. A mapping table of "data identifier ID-storage path-associated business data ID" is established in a MySQL database to enable the association and retrieval of unstructured and structured data. Unstructured data storage supports breakpoint resume and boasts a high upload success rate. ; Cold data archiving is used for historical data that has exceeded its retention period (such as raw data from high-frequency sensors from 2 years ago, or calibration records from decommissioned dummies). It employs a "compressed archiving + offline storage" technique, first compressing the data using the GZIP compression algorithm (compression ratio...). Then, the data is migrated to a low-cost SATA hard drive array-based cold storage node. The cold storage node supports periodic wake-ups to verify data integrity, ensuring the recoverability of archived data (data recovery success rate). This reduces the capacity pressure on the main storage node.

[0030] To support the time series data construction module in quickly retrieving data by "dummy ID-component ID-model type", a composite index is created in the MySQL database. Specifically, a joint index of "dummy_id + component_id + model_type" is designed for the "preprocessed data table". Tree index structures optimize query efficiency, reducing retrieval time by three dimensions from seconds without an index to the level of seconds. Shortened to millisecond level Meanwhile, a tag index is created in the InfluxDB time series database for "sensor_id + time_range", which supports quick filtering of high-frequency time series data by sensor ID (associated component ID) and time range, meeting the high-efficiency data access requirements of the time series construction module.

[0031] In the time series construction module, the preprocessed multi-source dummy data is constructed into a time series dataset according to dummy ID and component ID as follows: Step 1: Based on the business scenario of dummy calibration and the input requirements of multi-model fusion algorithms, extract basic identification features, working condition quantitative features, degradation trend features, and environmental correlation features from the preprocessed multi-source dummy data. Among them, the basic identification features include dummy ID, component ID, calibration equipment serial number, and calibration method version number (standardized according to "method number-version number"). String encoding is used to ensure uniqueness and adapt to the model's requirement to distinguish individual samples. The quantification of working condition characteristics includes quantifying non-numerical / discrete data such as impact energy, point of impact, and test type, using the formula: "Working condition strength coefficient = (Impact energy / Component rated withstand energy)". The comprehensive working condition index is calculated by weighting the test type (1.0 for whole vehicle collision, 0.7 for partial collision, and 0.5 for bench test). At the same time, the points of action are converted into numerical labels according to the body parts of the dummy (such as chest C1-C3, legs L1-L2) to adapt the quantitative analysis of the impact of the model on the working condition. The degradation trend feature is based on dynamic features extracted from time series windows, including "mean / variance of the last 3 calibration deviations, rate of change of calibration deviation (current deviation - historical mean) / historical mean, and percentage of cumulative usage (current usage / historical average failures)", which directly depicts the degradation pattern of components and strengthens the trend correlation of time series data; The environmental correlation feature transforms environmental data such as temperature and humidity into "Environmental Deviation Coefficient = (Current Environmental Value - Standard Environmental Value) / Standard Environmental Value" (standard environmental value is calculated as follows). Standard setting is temperature ,humidity It also adds an environmental stability index, namely the standard deviation of the environmental deviation coefficient after three consecutive calibrations, to quantify the impact of environmental fluctuations on sensor performance and adapt the model to the modeling requirements of environmental interference factors.

[0032] Step 2: Based on the baseline time axis anchoring and data interpolation completion, perform time-series alignment and feature data fusion processing on the multi-source dummy data to generate fused time-series data; specifically: First, the time of the calibrated event is used as the core anchor point, and a baseline timeline is constructed according to the preset minimum time unit. For example, the baseline timeline is constructed with the day as the minimum time unit. If there are multiple calibrations on a single day, the anchor points are refined by hour-minute-second to ensure that each time node corresponds uniquely. Next, the collected dummy multi-source data is divided into high-frequency sensor data, low-frequency data, and event-triggered data. The high-frequency sensor data is aggregated according to the calibration event time window, such as from 1 hour before to 1 hour after calibration; and associated with the low-frequency data of the calibration event, such as calibration deviation and operating parameters; the event-triggered data (such as component maintenance and standard change) is anchored to the reference time axis according to the occurrence time axis and marked as event labels (such as maintenance label 1, no maintenance label 0). Finally, the basic identification features, working condition quantification features, decay trend features, and environmental correlation features of the same dummy component are concatenated in chronological order into a high-dimensional time-series vector to represent the fused time-series data. The vector dimension is fixed at 28 dimensions to ensure the uniformity of the input dimensions of multiple models. An example of the generated time-series dataset is shown below: Example 1: Dummy ID: ZPZ408 Component: Ribs Calibration Date: 2025-08-15 Number of uses: 42 Initial chest displacement: 104.5 mm Chest displacement calibration result: 35.1 mm Calibration deviation:

[0033] Ambient temperature: Ambient humidity: Operator: Zhang San.

[0034] Example 2: Dummy ID: ZPZ409 Component: Ribs Calibration Date: 2025-09-28 Number of uses: 44 Initial chest displacement: 104.8 Chest displacement calibration result: 35.1 mm Calibration deviation:

[0035] Ambient temperature: Ambient humidity: Operator: Li Si.

[0036] Step 3: Based on the input characteristics of each model in the multi-model fusion algorithm, perform differential optimization on the fused time-series data to generate multiple model-specific datasets. In this embodiment, the models include statistical models, machine learning models, and deep learning models. Statistical models include, for example, the Weibull model and the Cox model. Data differential optimization for statistical models includes normalizing numerical features (such as calibration deviation and operating condition intensity coefficient) using Z-score standardization to ensure that the data distribution conforms to the model assumptions. Machine learning models include, for example, the XGBoost model. Data differential optimization for machine learning models includes one-hot encoding of discrete features (such as test type labels and event labels) and encoding of highly correlated features (such as ambient temperature and humidity, correlation coefficients). Principal component analysis (PCA) was used for dimensionality reduction to retain... While reducing feature redundancy, this improves model training efficiency. For deep learning models such as the LSTM-Transformer, data discrepancy optimization includes standardizing the sequence length of the time-series vectors, truncating the sequence based on the "most recent 30 time nodes" (padding with forward padded if less than 30), and normalizing the feature values. The interval is adapted to the requirements of deep learning models on the length and numerical range of the input sequence; at the same time, "temporal position encoding features" are extracted separately for the Transformer model to mark the relative position of each time node in the sequence, thereby strengthening the model's ability to capture temporal dependencies.

[0037] Step 4: Embed version metadata for each model-specific dataset, establish a mapping relationship with the model version, and store it in a database with a 3D index including dummy ID, component ID, and model type. Version metadata embedding involves embedding "calibration standard version number and feature processing rule version number" into each record of the time-series dataset. The feature processing rule version number is associated with specific processing logic, such as V1.0 corresponding to Z-score standardization and V2.0 corresponding to Min-Max normalization. When the AI ​​model calls the data later, the system automatically matches the corresponding version of the time-series data according to the version number bound during model training. If the data version is inconsistent with the model requirements, a version conversion is automatically triggered, such as reprocessing V2.0 version data according to V1.0 rules to avoid model prediction bias caused by version differences.

[0038] The core layer receives time-series datasets, models the performance degradation of dummies and components, and predicts their remaining usable life based on a multi-model fusion algorithm, outputting the prediction results. The core layer includes a model library module, a training module, an inference module, a fusion module, and an interpretation module. The model library module integrates statistical, machine learning, and deep learning models for predicting the lifespan of dummy components. It automatically selects models based on the data characteristics of time-series datasets and allows for manual model selection. In this embodiment, the parameters and adaptation ranges of different models are shown in the table below:

[0039] Each of the above models is bound to a unique version number in the format of model type-version number-training date, such as "Weibull-V1-20250815", and associated with the training data version and feature engineering rule version to ensure that the data source and processing logic can be traced back when the model is called.

[0040] The training module is used to extract data features from the time series dataset called from the data layer, load the target model from the model library module, iterate the training using the historical data training set with a preset optimizer and loss function, calculate the model evaluation index using the historical data validation set, and generate a standardized model package. In this embodiment, the time series dataset is first divided into k consecutive windows according to time order, with the first k-1 windows serving as the training set and the kth window as the validation set. Then, a hybrid loss function is constructed for the RUL prediction task (remaining useful life), with the expression:

[0041] in This is the weighting coefficient, with a default value of 0.6. Mean square error, Mean absolute error; Subsequently, a dedicated optimizer was configured for different model types. For statistical models, maximum likelihood estimation and partial likelihood estimation were used to solve for the parameters, and the iteration termination condition was the change in parameters. or number of iterations The machine learning model uses grid search to optimize hyperparameters, with a search range including tree depth of 3-8, learning rate of 0.01-0.1, and 100-500 iterations. The objective function is to minimize the validation set RMSE. The deep learning model uses the Adam optimizer with an initial learning rate of 0.001. The learning rate decay strategy is to decrease to 0.9 times the original rate every 50 epochs, and the weight decay coefficient is... To prevent overfitting; After training, a standardized model package is generated, containing model files in .pkl / .h5 format; data processing rules such as feature normalization parameters and outlier removal thresholds; and evaluation reports: training set / validation set MAE / RMSE. Training logs include iteration counts and loss function change curves, ensuring the model is portable and reproducible.

[0042] The inference module receives the time-series dataset from the data layer, calls the corresponding standardized model package trained by the training module to perform inference on the time-series dataset, and outputs the RUL prediction results, failure probability curves, and model version identifier. In this embodiment, the inference module's inference modes include online inference and offline inference. Online inference provides real-time services based on the REST / gRPC interface, supporting RUL prediction for one dummy ID-part ID per request, with a response time of [missing information]. Batch request support Parallel computation of dummy ID and component ID, response time The scenario where RUL is obtained in real time after adaptation testing; Offline inference is based on Airflow task scheduling. It automatically reads the newly added calibration data from the previous day at midnight every day, calculates the RUL of all dummy parts in batches, stores the results in the result library and updates the visualization interface, and is suitable for scenarios with periodic full prediction.

[0043] During inference, the input format is JSON, which must include dummy ID (string), component ID (string), latest feature set (such as cumulative usage count, last 3 calibration deviations, average impact energy), and prediction type (RUL / failure probability curve). The output format is JSON, returning RUL point values, confidence intervals, failure probability curve data, model version, and inference time.

[0044] The fusion module selects multiple optimal models based on the evaluation accuracy in the training module, and then uses the latest validation set of the optimal models. After calculating the weights of each model, RUL prediction and prediction results fusion are performed separately to obtain a fused model package. In this embodiment, a weighted voting method is used to fuse the prediction results of the best model in the model library module, for example, three best models. The weight calculation is based on the model's performance on the validation set. Values, expressed as: Weights = Model Value / All Candidate Models The sum of the values ​​ensures that the model with higher accuracy has a larger weight; then, the three models are called to predict the RUL for the same input feature, and the predicted values ​​of each model are multiplied by their corresponding weights and summed to obtain the fused RUL point values. The fused model corresponding to this fusion result is then packaged into a model package.

[0045] The uncertainty module is used to resample the input time series dataset with replacement, generate a new feature dataset, and perform RUL prediction on the new feature dataset based on the fused model package. Confidence intervals are extracted from the output RUL prediction values, and the confidence intervals are validated. In this embodiment, the uncertainty estimation method uses the Bootstrap resampling method. Confidence interval, with specific parameters: number of resampling attempts = 100, confidence level = By resampling multiple times to simulate the impact of data fluctuations on the prediction results, we can ensure that the confidence interval can cover the probability of the true RUL. Specifically: The input time-series features are resampled with replacement to generate 100 new feature datasets; the fused model package is used to perform RUL prediction on each of the 100 resampled datasets, resulting in 100 RUL prediction values; the 100 RUL prediction values ​​are sorted in ascending order, and the 2.5th percentile (3rd value) is taken as the lower limit of the confidence interval, and the 97.5th percentile (98th value) is taken as the upper limit of the confidence interval; if the confidence interval width is (upper limit - lower limit)... If the model is used for the first time, it will be determined that the prediction uncertainty is too high, and the model will be automatically retrained (by increasing the amount of training data). The output results will also indicate that the calibration frequency should be increased to reduce the uncertainty.

[0046] The inference module is also used to adjust the called model based on the validation results; for example, to re-select the model.

[0047] The explanation module is used to calculate the influence of each feature on the RUL prediction result based on the RUL prediction result output by the inference module and the corresponding input features, and to select several key features and generate an adjustment scheme for each key feature; specifically: Obtain the model version identifier, RUL prediction results, and input feature dataset transmitted by the inference module; The model type is identified based on the model version identifier, and different preset feature influence value calculation methods are invoked to calculate the corresponding feature influence values ​​according to different model types; specifically: If the model is a statistical model, the characteristic influence value is calculated based on the parameter influence coefficient method, and the expression is:

[0048] in, For the first The model coefficients for each feature, For the first The range of values ​​for each feature; For the statistical model corresponding to the first The influence value of each feature; If the model is a machine learning model, the feature influence value is calculated based on the feature gain method, and the expression is:

[0049] in, This represents the influence value of the i-th feature in the machine learning model. The total information gain of the i-th feature is... The sum of the total information gains of the n features; If the model is a deep learning model, the feature influence value is calculated based on the SHAP value method, and the expression is:

[0050] in, For SHAP value mapping functions, Let be the range of values ​​for the i-th feature; This represents the influence value of the i-th feature in the deep learning model.

[0051] The characteristic influence values ​​calculated by different methods are normalized to The intervals are then divided into levels according to preset thresholds, associated with business names, and a feature importance ranking table is output; an example is shown below:

[0052] The explanation is as follows: The cumulative number of uses is the core factor affecting the rib cage RUL of the dummy, contributing a certain percentage. For every 10 additional uses, the RUL decreases by an average of 3 times, and this indicator threshold warning should be given priority attention.

[0053] Based on a feature importance ranking table, the top N key impact features are extracted (e.g., the top 3). Business constraint rules are preset, and each key impact feature is adjusted within a preset adjustment range to generate candidate adjustment values. Simultaneously, values ​​violating the business constraint rules are removed, and the remaining candidate adjustment values ​​are retained. The expression is:

[0054] in, Candidate adjustment values, , , To define the range of feature values ​​allowed by the preset business constraint rules, Let be the range of values ​​for the i-th feature. To pre-determine the adjustment range, this application is... , indicating based on the current value Adjust and generate candidate adjustment values.

[0055] Keeping the feature values ​​of the remaining features unchanged, we replace the feature values ​​of the key features with candidate adjustment values ​​to construct multiple candidate feature vectors. The inference module is invoked, the candidate feature vector is input into the model package of the original inference module, the corresponding RUL prediction result is output, and the RUL change is calculated. Based on a preset priority scoring function, the priority score is calculated according to the change in RUL, expressed as follows:

[0056] in, The change in RUL Rate the priority. To adjust the implementation cost of the solution, it is set according to the business scenario: low = 1, medium = 2, high = 3. For example, the cost of reducing impact energy is 1, and the cost of pre-calibration is 2.

[0057] Priority rating Sort in descending order and filter. The proposed scheme retains the top N priority schemes; this application retains the top 3, and the output example is as follows:

[0058] The explanation is as follows: The optimal solution is to reduce the working condition strength coefficient. The first option is to implement it at a low cost and extend the RUL's lifespan by three times, which can be achieved by optimizing the test conditions. The second-best option is to control the average calibration deviation, which requires enhancing the calibration accuracy during the next calibration.

[0059] The application layer receives the prediction results output by the core layer, automatically updates the dummy's status, pushes spare parts procurement requirements, and provides a visual display. The application layer includes a dummy and component status automatic update module, a spare parts requirement generation module, and a visualization module. The automatic status update module for dummies and components is used to construct a four-level status system (normal, early warning, emergency, and failure) and a built-in judgment algorithm based on the prediction results output by the core layer, and automatically update the status of dummies and components; examples of judgment are shown in the table below:

[0060] The spare parts demand generation module is used to construct a procurement demand calculation model based on the prediction results of the core layer and spare parts data, and output a standardized procurement demand form. In this embodiment, the system connects to the ERP system every morning to obtain the current inventory quantity and spare parts delivery date of each component. It compares the RUL prediction results with the demand trigger threshold to trigger procurement demand. Combining inventory data, delivery date and dummy usage plan, it calculates the suggested procurement quantity and the latest order placement time to generate a standardized procurement demand form, which includes component ID, component name, dummy matching model, suggested procurement quantity, latest order placement time, and demand basis (RUL prediction results, inventory status).

[0061] The visualization module is used to visualize the core layer prediction process, prediction results, the status of dummies and components, and standardized procurement requirements.

[0062] like Figure 2 As shown, in another embodiment of this example, a calibration lifecycle prediction method for crash test dummies is also included, applied to the aforementioned calibration lifecycle prediction system for crash test dummies, comprising: S1: Receive full-scene data for dummy calibration, preprocess and store it to generate preprocessed multi-source dummy data; S2: Construct a time series dataset from the preprocessed multi-source dummy data according to dummy ID and component ID; S3: Input the time series dataset into the multi-model fusion algorithm to model the performance degradation of the dummy and its components and predict the remaining usable life, and output the prediction results; S4: Based on the output prediction results, automatically update the dummy status, push spare parts procurement requirements, and display them visually.

[0063] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A calibration lifecycle prediction system for crash test dummies, characterized in that: It includes the data source layer, data layer, core layer, and application layer, among which: The data source layer serves as the data input source, covering data from the entire scenario of dummy calibration. The data layer includes a data acquisition and storage module and a time series construction module. The data acquisition and storage module is used to access the data source layer, collect multi-source data from dummies, and store it after preprocessing. The time series construction module is used to construct a time series dataset from the preprocessed multi-source data of dummies according to the dummy ID and component ID. The core layer receives time-series datasets, models the performance degradation of dummies and components, and predicts their remaining usable life based on a multi-model fusion algorithm, then outputs the prediction results. The application layer is used to receive the prediction results output by the core layer, automatically update the dummy status, push spare parts procurement requirements, and display them visually. Specifically, the time series construction module constructs a time series dataset from the preprocessed multi-source dummy data according to dummy ID and component ID as follows: Based on the business scenario of dummy calibration and the input requirements of multi-model fusion algorithm, basic identification features, working condition quantification features, decay trend features and environmental correlation features are extracted from preprocessed multi-source dummy data. Based on benchmark time axis anchoring and data interpolation completion, time-series alignment and feature data fusion processing are performed on multi-source dummy data to generate fused time-series data. Based on the input characteristics of each model in the multi-model fusion algorithm, the fused time series data is optimized for differences, generating multiple sets of model-specific datasets; Version metadata is embedded in each model's unique dataset, and a mapping relationship is established with the model version. This metadata is then stored in a database containing a 3D index that includes dummy ID, component ID, and model type.

2. The calibration life-cycle prediction system for crash test dummies according to claim 1, characterized in that: The process of performing time-series alignment and feature data fusion on multi-source dummy data based on reference time axis anchoring and data interpolation completion to generate fused time-series data is as follows: Using the time of the calibrated event as the core anchor point, a baseline timeline is constructed according to a preset minimum time unit; The collected dummy multi-source data is divided into high-frequency sensor data, low-frequency data, and event-triggered data. The high-frequency sensor data is aggregated according to the calibration event time window and associated with the low-frequency data of the calibration event. The event-triggered data is anchored to the reference time axis according to the occurrence time axis and marked as event labels. The basic identification features, working condition quantification features, decay trend features, and environmental correlation features of the same dummy component are concatenated in chronological order to form a high-dimensional time-series vector, which represents the fused time-series data.

3. The calibration life-cycle prediction system for crash test dummies according to claim 2, characterized in that: The core layer includes a model library module, a training module, an inference module, a fusion module, an uncertainty module, and an interpretation module, wherein: The model library module is used to integrate statistical models, machine learning models, and deep learning models for predicting the lifespan of dummy parts, and to automatically select models based on the data characteristics of time series datasets and to manually specify the models to select. The training module is used to extract data features from the time series dataset called from the data layer, load the target model in the model library module, iterate the training using the training set of historical data according to the preset optimizer and loss function, calculate the model evaluation index using the validation set of historical data, and generate a standardized model package. The inference module receives the time series dataset from the data layer, calls the corresponding standardized model package trained by the training module to perform inference on the time series dataset, and outputs the RUL prediction results, failure probability curves and model version identifier. The fusion module selects multiple optimal models based on the evaluation accuracy in the training module, and then uses the latest validation set of the optimal models. After calculating the weights of each model, RUL prediction and prediction results are fused to obtain the fused model package. The uncertainty module is used to resample the input time series dataset with replacement to generate a new feature dataset, and to perform RUL prediction on the new feature dataset according to the fused model package. The confidence intervals of the output RUL prediction values ​​are extracted and the confidence intervals are verified. The inference module is also used to adjust the called model based on the verification results; The explanation module is used to calculate the influence of each feature on the RUL prediction result based on the RUL prediction result output by the inference module and the corresponding input features, and to select several key features and generate an adjustment scheme for each key feature.

4. The calibration life-cycle prediction system for crash test dummies according to claim 3, characterized in that: The uncertainty module resamples the input time series dataset with replacement to generate a new feature dataset. It then performs RUL prediction on the new feature dataset based on the fused model package. Confidence intervals are extracted from the output RUL prediction values, and the confidence intervals are validated as follows: The time series features in the input time series dataset are resampled with replacement to generate a new feature dataset. The fused model package is called to perform RUL prediction on the new feature dataset, and new RUL prediction results are obtained. Sort the new RUL prediction results from smallest to largest, take the 2.5th percentile as the lower limit of the confidence interval, and the 97.5th percentile as the upper limit of the confidence interval; The confidence interval is validated for its width, and the validation result is fed back to the inference module to help the inference model optimize its calling strategy.

5. The calibration life-cycle prediction system for crash test dummies according to claim 4, characterized in that: The explanation module calculates the influence of each feature on the RUL prediction result based on the RUL prediction result output by the inference module and the corresponding input features, and selects several key features. For each key feature, an adjustment scheme is generated as follows: Obtain the model version identifier, RUL prediction results, and input feature dataset transmitted by the inference module; The model type is identified by the model version identifier, and different preset feature influence value calculation methods are called to calculate the corresponding feature influence values ​​according to different model types. The characteristic influence values ​​calculated by different methods are normalized to The intervals are divided into levels according to preset thresholds, associated with business names, and a feature importance ranking table is output. Based on the feature importance ranking table, the top N key impact features are extracted. Pre-defined business constraint rules are then applied to each key impact feature according to a preset adjustment range to generate candidate adjustment values. Simultaneously, values ​​violating the business constraint rules are removed, and the remaining candidate adjustment values ​​are retained. The expression is as follows: in, Candidate adjustment values, , , To define the range of feature values ​​allowed by the preset business constraint rules, Let be the range of values ​​for the i-th feature. Preset adjustment range; Keeping the feature values ​​of the remaining features unchanged, we replace the feature values ​​of the key features with candidate adjustment values ​​to construct multiple candidate feature vectors. The inference module is invoked, the candidate feature vector is input into the model package of the original inference module, the corresponding RUL prediction result is output, and the RUL change is calculated. Based on a preset priority scoring function, the priority score is calculated according to the change in RUL, expressed as follows: in, The change in RUL Rate the priority. To adjust the implementation cost of the solution, it is set according to the business scenario; Priority rating Sort in descending order and filter. The solution is to retain the top N priority schemes.

6. The calibration life-cycle prediction system for crash test dummies according to claim 5, characterized in that: The model type is identified based on the model version identifier, and different preset feature influence value calculation methods are called to calculate the corresponding feature influence values ​​according to different model types. Specifically: If the model is a statistical model, the characteristic influence value is calculated based on the parameter influence coefficient method, and the expression is: in, For the first The model coefficients for each feature, For the first The range of values ​​for each feature; For the statistical model corresponding to the first The influence value of each feature; If the model is a machine learning model, the feature influence value is calculated based on the feature gain method, and the expression is: in, This represents the influence value of the i-th feature in the machine learning model. The total information gain of the i-th feature is... The sum of the total information gains of the n features; If the model is a deep learning model, the feature influence value is calculated based on the SHAP value method, and the expression is: in, For SHAP value mapping functions, Let be the range of values ​​for the i-th feature; This represents the influence value of the i-th feature in the deep learning model.

7. The calibration life-cycle prediction system for crash test dummies according to claim 1, characterized in that: The application layer includes a dummy and component status automatic update module, a spare parts requirement generation module, and a visualization display module, wherein: The automatic status update module for dummies and components is used to construct a four-level status system of normal, early warning, emergency, and failure based on the prediction results output by the core layer, and to automatically update the status of dummies and components. The spare parts requirement generation module is used to build a procurement requirement calculation model based on the prediction results of the core layer and spare parts data, and output a standardized procurement requirement sheet. The visualization module is used to visualize the core layer prediction process, prediction results, the status of dummies and components, and standardized procurement requirements.

8. A method for predicting the entire lifecycle of a crash test dummy during calibration, applied to the system for predicting the entire lifecycle of a crash test dummy during calibration as described in any one of claims 1-7, characterized in that: include: S1: Receive full-scene data for dummy calibration, preprocess and store it to generate preprocessed multi-source dummy data; S2: Construct a time series dataset from the preprocessed multi-source dummy data according to dummy ID and component ID; S3: Input the time series dataset into the multi-model fusion algorithm to model the performance degradation of the dummy and its components and predict the remaining usable life, and output the prediction results; S4: Based on the output prediction results, automatically update the dummy status, push spare parts procurement requirements, and display them visually.

Images