Dummy sensor abnormity intelligent diagnosis system and method

Abstract

The invention belongs to the technical field of dummy sensor diagnosis, and particularly relates to a dummy sensor abnormity intelligent diagnosis system, which establishes a dummy and sensor unique identity index through a data acquisition module, and acquires and associates environmental test data and sensor data; the preprocessing module completes data cleaning, impact section interception and feature vector generation and normalization; the anomaly detection module generates a corresponding result through multi-dimensional detection of a rule layer, a statistical layer and a model layer, and the intelligent comparison module outputs a related score and a result through comparison of individual historical data and group data; the exception scoring module is fused with the multi-source scores to calculate a unified exception score; the dynamic variable analysis module locates fault causes through multi-step analysis, and the fault code generation module is matched with a fault code system to output fault codes, cause instructions and processing suggestions. According to the invention, the problem that the existing dummy sensor anomaly diagnosis technology is lack of multi-dimensional data association analysis capability can be solved.

Description

Technical Field

[0001] This invention belongs to the field of dummy sensor diagnostic technology, and particularly relates to an intelligent diagnostic system and method for dummy sensor anomalies. Background Technology

[0002] In the field of automotive safety performance evaluation, crash testing is a core step in verifying a vehicle's crash protection capabilities and ensuring occupant safety. Crash dummies, as key equipment simulating the force and motion of the human body during a collision, are equipped with multi-axis force / accelerometers covering core areas such as the head, torso, pelvis, and knees. These dummies are crucial components for accurately collecting occupant force parameters and acceleration changes during a collision. This data directly determines the scientific validity and reliability of the crash test results, thereby influencing the direction of vehicle safety design optimization and the final safety performance rating.

[0003] However, in actual testing scenarios, dummy sensors are susceptible to various abnormal phenomena due to multiple factors. First, environmental variables such as temperature fluctuations, humidity changes, and electromagnetic radiation in the collision test environment directly affect the stability of the sensor's circuit and the accuracy of signal transmission. Second, during sensor installation, problems such as misalignment of the installation position, insufficient or excessive fixing torque, or poor contact of the connecting harness can disrupt the rigid connection between the sensor and the dummy's body structure, causing the sensor to fail to accurately follow the dummy's movement during the collision, resulting in sudden data changes and abnormal peak values. Third, as the number of times the dummy is used increases, the sensor's internal elastic elements, strain gauges, and other core components will experience fatigue wear after repeated collisions and impacts, gradually degrading its performance. At the same time, the sensor's calibration coefficients will naturally drift over time, causing the sensor's output data to deviate from the standard range in the absence of external interference, i.e., "inaccuracy."

[0004] To address the aforementioned sensor anomaly issues, existing technologies primarily employ manual inspection or simple threshold determination. However, these methods have significant drawbacks and fail to meet the requirements of crash tests for data reliability and processing efficiency. Firstly, manual inspection relies on engineers' experience, requiring the verification of historical sensor data and test records one by one. This is not only time-consuming and labor-intensive but also fails to identify subtle parameter drifts or latent anomalies, resulting in insufficient accuracy and comprehensiveness in anomaly detection. Secondly, the simple threshold method only determines whether a sensor is abnormal by setting a fixed data range, without considering the influence of dynamic variables such as test conditions, environmental parameters, and the sensor's service life.

[0005] In summary, existing dummy sensor anomaly detection technologies lack the ability to perform correlation analysis on multi-dimensional data, intelligent anomaly identification methods, and efficient processing mechanisms. As a result, they are unable to quickly and accurately detect sensor anomalies and locate the causes of anomalies. This leads to abnormal data potentially affecting the reliability of crash test results and even threatening dummy safety. At the same time, traditional technologies cannot systematically manage sensor maintenance, which can easily lead to calibration delays and further increase the risk of sensor anomalies. Summary of the Invention

[0006] The technical problem solved by this invention is to provide an intelligent diagnostic system and method for abnormal dummy sensors, so as to solve the problem that existing abnormal dummy sensor diagnostic technologies lack the ability to perform multi-dimensional data correlation analysis.

[0007] The basic solution provided by this invention is a dummy sensor anomaly intelligent diagnostic system, comprising a data acquisition module, a preprocessing module, an anomaly detection module, an intelligent comparison module, an anomaly scoring module, a dynamic variable analysis module, and a fault code generation module, wherein: The data acquisition module is used to collect environmental test data and sensor data of the dummies and sensors after establishing a unique identification index for each dummy and its corresponding sensor; The preprocessing module is used to preprocess the sensor data, extract the impact segment data according to a fixed window length, extract key features from the impact segment data, and generate feature vectors; the preprocessing module also performs normalization processing on the environmental test data; The anomaly detection module determines hardware faults in normalized environmental test data based on the rule layer, calculates anomaly scores for sensor data anomalies based on the statistics layer, and calculates anomaly scores for feature vectors based on the model layer, generating hardware fault results, statistical anomaly scores, and model anomaly scores. The intelligent comparison module is used to compare historical data of the same sensor and the same sensor position of the same type of dummy, and generate individual similarity scores and group anomaly results respectively. The anomaly scoring module is used to integrate statistical anomaly scores, model anomaly scores, and individual similarity scores to calculate a unified anomaly score; The dynamic variable analysis module is used to receive unified anomaly scores, generate target response quantities, extract key variables to be verified, perform prior screening, nearest neighbor matching, regression verification and sensitivity analysis, and output the causes of failures. The fault code generation module is used to match a preset fault code system based on the fault cause, hardware fault result, and group anomaly result, and output the fault code, cause description, and handling suggestions.

[0008] Furthermore, the unique identification index for each dummy and its corresponding sensor is established in the data acquisition module as follows: Each dummy is assigned a unique ID. Sensors are numbered according to dummy ID, sensor model, and installation location, and basic sensor information is registered. The system collects and stores environmental test data, sensor data, and historical data from the crash test dummy. It also establishes a cross-data type association index using dummy ID, sensor number, and test time as core key fields to generate a multi-dimensional data association map.

[0009] Furthermore, the anomaly detection module includes a rule layer unit, a statistical layer unit, and a model layer unit, wherein: The rule layer unit is used to receive preprocessed sensor data and directly determine hardware-level anomalies when there are constant values, NaN values, or over-range conditions, and output hardware fault results. The statistical layer unit calculates the standardized score of the feature vector based on the historical data distribution of sensors of the same model and location, and transforms it into anomaly scores. The expression is:

[0010]

[0011] in, These are abnormal scores. For the first The average of the standardized scores of the eigenvectors, For the first The threshold of the standardized scores of the feature vectors. For adjustment coefficients; For the first Standardized scores of each feature vector. For the first The values ​​of each eigenvector, The first in historical normal data The mean of the eigenvectors, The first in historical normal data The standard deviation of each eigenvector; The statistical layer unit also tracks the drift trend of the feature vector based on an exponentially weighted moving average and calculates the trend anomaly score of the feature vector. ; and based on the abnormal score and trend anomaly score Perform weighted summation to generate statistical outlier scores. The expression is:

[0012]

[0013]

[0014] in, For trend anomaly scores, This indicates a cutoff function to prevent over-range drift from rendering the fractions meaningless; For the first The exponentially weighted moving average of the sensor target feature vector in this experiment. For the first The mean of historical normal data for each feature vector; These are the weighting coefficients. For the first The feature vector values ​​of this trial For the first The exponentially weighted moving average of the trials; Weights for outlier scores. Weights for trend anomaly scores; The model layer unit is used to train the autoencoder model based on historical sensor data. It inputs the feature vector into the autoencoder model, outputs a reconstructed vector, and calculates the reconstruction error, expressed as:

[0015] Normalization to :

[0016] in, For reconstruction error, The number of feature vectors, These are the original values ​​of the feature vector. The reconstructed values ​​of the feature vector, For model anomaly scores, This represents the truncation function. This represents the minimum reconstruction error of the feature vector within historical normal data. This represents the maximum reconstruction error of the feature vector in historical normal data.

[0017] Furthermore, the intelligent comparison module includes an individual comparison unit and a group comparison unit, wherein: The individual comparison unit is used to compare historical data from the same sensor, calculate the DTW distance between the current waveform and the most recent N normal waveforms, and convert it into a similarity score. ; The group comparison unit is used to compare the feature vectors of sensors at the same position on the same type of dummy and calculate the cosine similarity. If the cosine similarity is lower than the preset threshold, it is marked as an abnormal individual in the group. If more than M sensors have a cosine similarity lower than the preset threshold, it is marked as a system abnormality and a group abnormality result is generated.

[0018] Furthermore, the dynamic variable analysis module includes a variable definition unit, a priori screening unit, a nearest neighbor matching unit, a regression validation unit, and a sensitivity analysis unit, wherein: The variable definition unit is used to select a uniform anomaly score as an indicator to measure the degree of anomaly, generate the target response quantity, and extract candidate anomaly variables from environmental test data. The prior screening unit is used to calculate the correlation between the abnormal candidate variables and the target response, screen out the TOP-K variables, and output a set of high-potential variables. The nearest neighbor matching unit is used to perform propensity score matching calculations for individual high-potential variables in the set of high-potential variables as a treatment group, and to set up a control group based on the propensity score of the treatment group. The average treatment effect of the treatment group and the control group is calculated, and the potential principal causes of the abnormality are output. The regression validation unit uses the uniform anomaly score as the dependent variable and the high-potential variable as the independent variable to construct a multiple linear regression model to calculate the variance inflation factor of each independent variable and output robust principal causal candidates for causal effects. The sensitivity analysis unit is used to quantify the relative contribution of high-potential variables to the unified anomaly score and output the main or compound causal factors as the causes of failure.

[0019] Furthermore, in the nearest neighbor matching unit, propensity score matching is performed on individual high-potential variables in the high-potential variable set to serve as the treatment group, and a control group is set up based on the propensity scores of the treatment groups. The average treatment effect of the treatment group and the control group is calculated, and the potential main causes of the anomaly are output as follows: Construct a logistic regression model with the single high-potential variable (whether the user belongs to the treatment group) as the dependent variable and other high-potential variables as independent variables. Calculate the propensity score for each dependent variable, expressed as follows:

[0020] in, For the first The propensity score of each dependent variable. For conditional probability, when the dependent variable When fixed, the probability of belonging to the treatment group is... This is the intercept term for logistic regression. Represents the regression coefficient of the dependent variable. Indicates the first p dependent variable values ​​for a sample; Treatment groups are generated based on propensity scores. For each high-potential variable in the treatment group, a corresponding control group is matched, where the control group has the smallest difference in propensity score compared to the treatment group. Based on the matched samples, the average treatment effect between the treatment group and the control group was calculated. The expression is:

[0021] in, This represents the number of samples in the processing group after nearest neighbor matching. This represents the number of samples in the control group after nearest neighbor matching; This represents the outlier response values ​​of all processed group samples after matching. Sum of abnormal response values To standardize anomaly scoring; Abnormal response values ​​of all control group samples after matching sum; Calculate the average treatment effect Standard error ,pass Test the calculation of average treatment effect of The value, expressed as:

[0022] according to Value lookup The distribution table determines the degrees of freedom, and the values ​​can be looked up based on the degrees of freedom. Value, comparison Is the value less than the significance level? If so, this indicates that the sample is a potential primary cause of the anomaly.

[0023] Furthermore, in the regression validation unit, a multiple linear regression model is constructed using a uniform anomaly score as the dependent variable and high-potential variables as independent variables to calculate the variance inflation factor of each independent variable, and the robust principal causal candidates for the causal effect are output as follows: To standardize anomaly scoring Using the variable with high potential as the dependent variable and the variable with high potential as the independent variable, a multiple linear regression model is constructed, with the expression as follows:

[0024] in, To standardize anomaly scoring, For the intercept term, The regression coefficients of the current independent variable are... For the current independent variable, For the first The regression coefficients of the independent variables, For the first One independent variable, This is the error term; Calculate the variance inflation factor between the individual independent variable and other independent variables. The expression is:

[0025] in, Let be the square of the multiple correlation coefficient between the k-th independent variable and the other independent variables. Let be the variance inflation factor of the k-th independent variable; Determine whether the variance inflation factor exceeds the preset collinearity threshold. If it does, remove the independent variable; otherwise, retain the independent variable and output the candidate set of main causes.

[0026] Furthermore, in the sensitivity analysis unit, the relative contribution of high-potential variables to the unified anomaly score is quantified, and the main or compound causal factors are output as the causes of the failure. The coefficients of each independent variable in the multiple linear regression model are converted into standardized regression coefficients, and the relative contribution of each independent variable to the uniform anomaly score is output, expressed as follows:

[0027] in, The standardized coefficient is... Let the standard deviation of the independent variable be . To standardize the standard deviation of outlier scores, The coefficients of the independent variables; the relative contribution is calculated by dividing the absolute value of the standardized coefficient of a single independent variable by the sum of the absolute values ​​of the standardized coefficients of all independent variables; The SHAP value quantifies the local contribution of each independent variable to a single sample and its global contribution to all samples. The expression for the local contribution is:

[0028] in, For the multiple linear regression model for the first The actual predicted anomaly score for each sample To remove the current variable The post-multiple linear regression model's prediction anomaly score for the i-th sample; This refers to the SHAP value of a single sample. The global contribution expression is:

[0029] in, The SHAP values ​​for the entire sample. The number of samples; The magnitude of a variable’s influence on anomalies is determined based on single-sample SHAP values ​​and full-sample SHAP values, and the cause of the failure is output.

[0030] Furthermore, it also includes a sensor calibration scheduling module, which dynamically adjusts the base cycle of abnormal sensor calibration scheduling based on usage frequency factors, environmental factors, and performance trend factors, and calls an intelligent scheduling model to generate calibration scheduling tasks based on the adjusted base cycle, abnormal sensor parameters, optimal targets, and constraints.

[0031] A method for intelligent diagnosis of dummy sensor anomalies, applied to the aforementioned intelligent diagnosis system for dummy sensor anomalies, includes: S1: After establishing a unique identification index for each dummy and its corresponding sensor, collect environmental test data and sensor data of the dummy and the sensor. S2: After preprocessing the sensor data, the impact segment data is truncated according to a fixed window length, and key features are extracted from the impact segment data to generate a feature vector; at the same time, the environmental test data is normalized. S3: Based on the rule layer, hardware faults are determined in the normalized environmental test data; based on the statistical layer, anomaly scores are calculated for sensor data anomalies; based on the model layer, anomaly scores are calculated for feature vectors to generate hardware fault results, statistical anomaly scores, and model anomaly scores. S4: Compare historical data from the same sensor and the same sensor on the same model of dummy to generate individual similarity scores and group anomaly results respectively; S5: Integrate statistical anomaly scores, model anomaly scores, and individual similarity scores to calculate a unified anomaly score; S6: Receives unified anomaly scores, generates target response quantities, extracts key variables to be verified, performs prior screening, nearest neighbor matching, regression verification and sensitivity analysis, and outputs fault causes; S7: Match the preset fault code system based on the fault cause, hardware fault result and group anomaly result, and output the fault code, cause description and handling suggestions.

[0032] The principles and advantages of this invention are as follows: The technical solution of this application takes full-link data association analysis + multi-dimensional anomaly verification + precise cause localization as its core principle. First, a unique identity index for the dummy and sensor is established through the data acquisition module to achieve precise binding of environmental test data and sensor data. Then, the preprocessing module completes data cleaning, impact segment interception and feature extraction to provide a standardized data foundation for subsequent analysis. The core of the system achieves preliminary screening of anomalies through multi-module collaboration: the anomaly detection module integrates rule-layer hardware fault judgment, statistical layer data anomaly scoring, and model-layer feature vector anomaly scoring; the intelligent comparison module combines individual historical data and group data of the same model to supplement the similarity verification dimension; the anomaly judgment module integrates multi-source anomaly scores to generate a unified anomaly score, providing a quantitative basis for cause analysis; finally, the dynamic variable analysis module locates the cause of the fault through a series of analyses such as prior screening and nearest neighbor matching, and the fault code generation module matches the preset system to output a complete diagnostic result.

[0033] Meanwhile, for the maintenance of abnormal sensors, a sensor calibration scheduling module is used to perform dynamic periodic calculation, intelligent scheduling optimization, and task tracking closed loop, which solves the problem of traditional "fixed periodic calibration and disordered resource allocation", reduces the risk of exceeding the time limit, and optimizes maintenance efficiency.

[0034] The advantages are as follows: It effectively solves the problem of existing dummy sensor anomaly diagnosis technology lacking multi-dimensional data correlation analysis capabilities. Through identity indexing, it achieves deep binding between environmental data and sensor data. Combining rule-based, statistical, and model-based three-layer detection with individual-group dual comparison, it constructs a comprehensive data correlation analysis system, significantly improving the comprehensiveness and accuracy of anomaly identification. Simultaneously, through dynamic variable analysis, it achieves precise tracing from anomaly detection to fault cause, avoiding the deficiency of traditional diagnosis which can only identify anomalies but cannot locate the root cause. Finally, it outputs standardized fault codes, cause explanations, and handling suggestions, realizing the automation and intelligence of the diagnostic process, significantly reducing the cost of manual intervention, and improving the reliability and diagnostic efficiency of dummy sensor test data. Attached Figure Description

[0035] Figure 1 This is a functional block diagram of an embodiment of the present invention; Figure 2 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0036] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1 As shown: A dummy sensor anomaly intelligent diagnostic system includes a data acquisition module, a preprocessing module, an anomaly detection module, an intelligent comparison module, an anomaly scoring module, a dynamic variable analysis module, and a fault code generation module, wherein: The data acquisition module is used to collect environmental test data and sensor data of the dummies and sensors after establishing a unique identification index for each dummy and its corresponding sensor; In this embodiment, each dummy is first assigned a unique ID. The sensors are then numbered according to the dummy ID, sensor model, and installation location. For example, HIII-ZPZ343-1 – Pelvis – 7264B-500 represents a dummy with ID HIII-ZPZ343-1, an installation location in the pelvis, and a sensor model 7264B-500. Basic sensor information is also registered, including the installation date: 2025-11-15; and the fixed torque, etc.

[0037] Subsequently, test environment data, sensor data, and historical data were collected and stored during the crash test on the dummy. For example, test environment data included collision speed, collision type, test temperature, humidity, electromagnetic interference intensity, etc.; sensor data included sensor output waveforms and waveform characteristic parameters (rise time, pulse width, etc.); and historical data included historical normal data, maintenance and replacement records, fault history, etc.

[0038] The collected data uses dummy ID, sensor number, and test time as core key fields to establish cross-data type association indexes and form a multi-dimensional data association map. For example, by using "dummy ID = HIII-ZPZ343-1", one can trace back all environmental data and sensor data of the dummy in all tests; by using "sensor number = pelvis = 7264B-500", one can summarize all output characteristics and calibration records of the sensor, realizing multi-dimensional data traceability.

[0039] The preprocessing module preprocesses the sensor data, extracts the impact segment data according to a fixed window length, and extracts key features from the impact segment data to generate a feature vector. The preprocessing module also normalizes the environmental test data. In this embodiment, the purpose of preprocessing is to eliminate interference factors in the data. First, the waveform data of the original sensor data undergoes DC filtering and bandpass filtering. DC filtering is used to remove static bias, and bandpass filtering is used to retain the effective frequency band of the collision impact. Then, the impact segment data is extracted according to a fixed window length, such as 10ms / window. The core features such as time domain and frequency domain are extracted from the impact segment data to form a feature vector. The time domain features include peak value, rise time, pulse width, AUC, and RMS. The frequency domain features include spectral centroid and energy proportion of a specific frequency band.

[0040] The normalization of environmental test data specifically involves normalizing data such as acceleration and force values ​​to a unified benchmark for different test conditions, ensuring comparability across test data.

[0041] The anomaly detection module performs hardware fault determination based on the rule layer of normalized environmental test data, calculates anomaly scores for sensor data anomalies based on the statistics layer, and calculates anomaly scores for feature vectors based on the model layer, generating hardware fault results, statistical anomaly scores, and model anomaly scores. The anomaly detection module includes rule layer units, statistical layer units, and model layer units, wherein: The rule layer unit is used to receive preprocessed sensor data and directly determine hardware-level anomalies when there are constant data values, NaN values, and over-range values, and output hardware fault results. In this embodiment, constant data value means that the sensor always outputs 0g, NaN value means that the sensor has no signal, and over-range means that the sensor exceeds its maximum range of 100g. These anomalies are directly marked as hardware-level anomalies and directly associated with subsequent fault codes.

[0042] The statistical layer unit calculates the standardized score of the feature vector based on the historical data distribution of sensors of the same model and location, and transforms it into anomaly scores. The expression is:

[0043]

[0044] in, These are abnormal scores. For the first The average of the standardized scores of the eigenvectors, For the first The threshold for the standardized scores of the eigenvectors is set to 2 in this application. To adjust the coefficient, this application uses a value of 0.5; For the first Standardized scores of each feature vector. For the first The values ​​of each eigenvector, The first in historical normal data The mean of the eigenvectors, The first in historical normal data The standard deviation of each eigenvector; Taking the HIII-ZPZ343-1 dummy pelvic sensor as an example, in one test acquisition, the peak value Z=2.33 (abnormal), pulse width Z=1.8 (normal), and RMS Z=2.1 (abnormal). Therefore, the three characteristics... The values ​​are 2.33, 1.8, and 2.1 respectively, with the abnormal features being the peak value and RMS. ; The current anomaly score is 0.61, which initially indicates a moderate degree of anomaly. The statistical layer unit also tracks the drift trend of the feature vector based on an exponentially weighted moving average and calculates the trend anomaly score of the feature vector. ; and based on the abnormal score and trend anomaly score Perform weighted summation to generate statistical outlier scores. The expression is:

[0045]

[0046]

[0047] in, This is the trend anomaly score, with a value range of [value range missing]. A higher score indicates more severe sensor drift and a more significant degree of anomaly. This indicates a cutoff function that ensures that when the drift exceeds... At this time, the maximum score is 1.0 to avoid drifting beyond the range and making the score meaningless; This represents the maximum allowable drift in this feature dimension, i.e., the critical threshold for sensor performance degradation. For example, the accuracy of an accelerometer is... , ,but ; For the first The exponentially weighted moving average of the sensor target feature vector in this experiment. For the first The mean of historical normal data for each feature vector is calculated from the historical normal data set of sensors of the same model and common installation location, such as the historical mean of the peak value of the HIII-ZPZ343-1 dummy pelvic sensor. ; This is a weighting coefficient, ranging from 0.1 to 0.3; in this application, it is set to 0.2. For the first The feature vector values ​​of this trial For the first The exponentially weighted moving average of the trials; Weights for outlier scores. Weights for trend anomaly scores; The model layer unit is used to train an autoencoder model based on historical sensor data. The autoencoder model consists of an encoder and a decoder. The encoder compresses high-dimensional sensor feature vectors into low-dimensional latent vectors and learns the core patterns of normal data. The decoder reconstructs the low-dimensional latent vectors back to the original feature dimensions, aiming to minimize the error between the reconstructed features and the original features. In this embodiment, the feature vectors are 7-dimensional feature vectors, such as the feature vectors extracted by the preprocessing described above. Therefore, the number of neurons in the encoder input layer is 7, and the output layer of the decoder layer is 7 Sigmoids. The remaining hidden layers, latent vector layers, loss functions, and optimizers are all conventionally set according to the input and output layers.

[0048] The feature vectors are input into the autoencoder model. These feature vectors are the seven-dimensional feature vectors extracted after the aforementioned sensor data preprocessing, including peak value, rise time, pulse width, AUC, RMS, spectral centroid, and energy percentage of a specific frequency band. The reconstructed vector is output, and the reconstruction error is calculated. The reconstruction error expression is then:

[0049] Normalization to :

[0050] in, For reconstruction error, The number of feature vectors is 7 in this application. These are the original values ​​of the feature vector. The reconstructed values ​​of the feature vector, The abnormality score of the model, with a value range of . , Indicates the truncation function, ensuring A value of 1.0 is used to mark a severe model anomaly, to avoid the score becoming meaningless due to over-range error. This represents the minimum reconstruction error of the feature vector within historical normal data. This represents the maximum reconstruction error of the feature vector in historical normal data.

[0051] The intelligent comparison module is used to compare historical data from the same sensor and the same sensor location on the same type of dummy, generating individual similarity scores and group anomaly results respectively. The intelligent comparison module includes an individual comparison unit and a group comparison unit, wherein: The individual comparison unit is used to compare historical data from the same sensor, calculate the DTW distance between the current waveform and the most recent N normal waveforms, and convert it into a similarity score. For example, the DTW distance between the current waveform of the HIII-ZPZ343-1 dummy pelvic position sensor and the three most recent normal waveforms is calculated. The smaller the distance, the higher the similarity, and this is converted into a similarity score. :

[0052] in, The similarity score for the current sensor. For DTW distance, For scale parameters; For example, DTW distance = 0.2 corresponds to a similarity score. The lower the score, the more significant the anomaly.

[0053] The group comparison unit is used to compare the feature vectors of sensors at the same position on the same dummy and calculate the cosine similarity. If the cosine similarity is lower than a preset threshold, it is marked as an abnormal individual in the group. If more than M sensors have a cosine similarity lower than the preset threshold, it is marked as a system anomaly and a group anomaly result is generated. For example, comparing the feature vectors of 10 HIII-ZPZ343-1 dummy pelvic position sensors of the same model and calculating the cosine similarity, if the similarity of the current sensor with the other 9 sensors is lower than 0.7, it is marked as an abnormal individual in the group. If more than 3 sensors have similar deviations, it is marked as a system anomaly.

[0054] The anomaly scoring module is used to integrate statistical anomaly scores, model anomaly scores, and individual similarity scores to calculate a unified anomaly score; specifically, the expression is:

[0055] in, To unify abnormal scores, To count outlier scores The weight is 0.3 in this application. For model anomaly scores The weight is 0.3 in this application. Similarity score The weight is 0.4; when the abnormal score is uniform. It is judged as a high-confidence anomaly when If it is determined to be suspicious and requires manual review, when It was determined to be normal.

[0056] The dynamic variable analysis module is used to receive unified anomaly scores and generate target response quantities. The module extracts key variables to be verified, performs prior screening, nearest neighbor matching, regression verification, and sensitivity analysis, and outputs the cause of the failure. The dynamic variable analysis module includes a variable definition unit, a prior screening unit, a nearest neighbor matching unit, a regression verification unit, and a sensitivity analysis unit. The variable definition unit is used to select a uniform anomaly score as an indicator to measure the degree of anomaly and generate the target response quantity. ; and extract outlier candidate variables from environmental test data; among which, the target response quantity The unified anomaly score was selected as a high-confidence anomaly (unified anomaly score will be discussed later). That is, the target response quantity Taking the abnormal diagnosis of the HIII-ZPZ343-1 dummy pelvic accelerometer sensor as an example scenario, candidate variables extracted from environmental test data include collision speed, ambient temperature, humidity, installation torque, vehicle model, and bench number, with the goal of locating the core variable that causes the sensor abnormality.

[0057] The prior screening unit is used to calculate the correlation between abnormal candidate variables and the target response, screen out the TOP-K variables, and output a set of high-potential variables. In this embodiment, the candidate variables are first converted into computable numerical variables. For example, for continuous variables such as ambient temperature, installation torque, collision data, and humidity, the original values ​​are used directly. For categorical variables, such as vehicle model and bench number, they are encoded as numerical values, such as vehicle model A=1, vehicle model B=2, etc. Then, the converted candidate variables are used to calculate the correlation index through a correlation algorithm. For linear correlations such as installation torque, the Pearson correlation coefficient is used. For nonlinear correlations such as humidity, the Spearman rank correlation is used. For categorical vectors, mutual information is used. In this way, the correlation index of each candidate variable is calculated and sorted. The TOP-3 variables with the largest absolute values ​​are selected as high-potential variables. For example, if the top 3 variables are installation torque, ambient temperature, and collision speed, then humidity, vehicle model, and bench number are excluded.

[0058] The nearest neighbor matching unit is used to perform propensity score matching calculations for individual high-potential variables in the set of high-potential variables, and to set up a control group based on the propensity scores of the treatment group. It calculates the average treatment effect between the treatment group and the control group, and outputs the potential principal causes of the anomaly; specifically: Construct a logistic regression model with the single high-potential variable (whether the user belongs to the treatment group) as the dependent variable and other high-potential variables as independent variables. Calculate the propensity score for each dependent variable, expressed as follows:

[0059] in, For the first The propensity score of each dependent variable. For conditional probability, when the dependent variable When fixed, the probability of belonging to the treatment group is... This is the intercept term for logistic regression. Represents the regression coefficient of the dependent variable. Indicates the first p dependent variable values ​​for a sample; For example, with "installation torque" as the variable T to be inspected, we define: Treatment group (T=1): Installation torque Test samples (below the standard threshold); Control group (T=0): Installation torque Test samples that meet the standards; Objective: To verify whether insufficient torque (T=1) leads to uniform anomaly scoring. Increase.

[0060] Subsequently, based on the constructed logistic regression model, with installation torque as the dependent variable and other high-potential variables such as temperature and speed as independent variables, the propensity score for each sample was calculated. Then, with temperature and speed as dependent variables and other high-potential variables as independent variables, the propensity scores were calculated separately. Treatment groups are generated based on propensity scores. For each high-potential variable in the treatment group, a corresponding control group is matched, where the control group has the smallest difference in propensity score compared to the treatment group. In this embodiment, an example of a treatment group and a control group is shown below:

[0061] Among them, S1's tendency scores were close to those of C1, C2, and C3, so the treatment group and the control group were successfully matched.

[0062] Based on the matched samples, the average treatment effect between the treatment group and the control group was calculated. The expression is:

[0063] in, This represents the number of samples in the processing group after nearest neighbor matching. This represents the number of samples in the control group after nearest neighbor matching; This represents the outlier response values ​​of all processed group samples after matching. Sum of abnormal response values To standardize anomaly scoring; Abnormal response values ​​of all control group samples after matching sum; In this embodiment, after matching the covariates of the treatment group and the control group, namely temperature and speed, there were no significant differences. The only difference was whether the installation torque was insufficient. At this point, the average treatment effect was calculated. The uniform anomaly score of sample S1 was 1.23, and the average uniform anomaly score Y of the control groups C1, C2, and C3 was 0.583. Therefore, the average treatment effect... ; Average treatment effect ,illustrate: Causal direction: Insufficient installation torque will cause the average sensor anomaly score to increase by 0.647; Effect strength: An increase exceeding the preset threshold of 0.4 is considered a significant effect.

[0064] Calculate the average treatment effect Standard error ,pass Test the calculation of average treatment effect of The value, expressed as:

[0065] according to Value lookup The distribution table determines the degrees of freedom, and the values ​​can be looked up based on the degrees of freedom. Value, comparison Is the value less than the significance level? If so, this indicates that the sample is a potential primary cause of the anomaly; where the above-mentioned treatment group and control group are single-sample scenarios, in reality it is a batch sample scenario, specifically 15 treatment groups and 45 control groups, therefore the calculation is... ,but The significance level When the critical value of t is 2.0, it is found that... This indicates that the variable has a causal effect on sensor anomalies, and is not a random fluctuation in the sample. Subsequently, a confidence interval is calculated using a 95% confidence level, and the result is substituted into... t critical value The final calculated 95% confidence interval is: Therefore, the above example outputs the following conclusion after the nearest neighbor matching unit: Statistical test results: Average treatment effect for "insufficient installation torque" and sensor malfunction. The standard error was calculated. , Statistic Corresponding (significance level) The null hypothesis was rejected, indicating that the causal effect of insufficient installation torque on sensor malfunction is real and significant. Confidence level conclusion: Based on Confidence level, calculated as follows The confidence interval is The fact that this interval does not include 0 further verifies the reliability of the above causal effect, indicating that the fluctuation range of the real effect is clear and there is no risk of distortion. The regression validation unit uses a uniform anomaly score as the dependent variable and high-potential variables as independent variables to construct a multiple linear regression model, calculates the variance inflation factor of each independent variable, and outputs robust principal causal candidates for causal effects; specifically: To standardize anomaly scoring Using the variable with high potential as the dependent variable and the variable with high potential as the independent variable, a multiple linear regression model is constructed, with the expression as follows:

[0066] in, To standardize anomaly scoring, For the intercept term, Let be the regression coefficient of the current independent variable. In this embodiment, based on the aforementioned example, the regression coefficient of torque is -0.052. For the current independent variable, For the first The regression coefficients for the independent variables are as follows: in this embodiment, the regression coefficient for temperature is 0.01, and the regression coefficient for speed is 0.005. For the first The independent variables also include interaction terms, such as the interaction between installation torque and temperature, and the interaction between installation torque and speed. The error term is used; the least squares method is used to fit the multiple linear regression model; the core objective of the multiple linear regression model is: 1. Verify whether the coefficients of the core independent variables are significant, i.e. 1. The sign is consistent with the nearest neighbor matching conclusion, confirming its robustness; 2. Verify whether other high-potential independent variables are significant, excluding the possibility that they are the main cause; 3. Verify whether the interaction terms of independent variables are significant, and determine whether there is a synergistic effect of variables; 4. Verify whether there is collinearity among variables, and ensure that the model results are reliable and free from interference.

[0067] Calculate the variance inflation factor between the individual independent variable and other independent variables. The expression is:

[0068] in, Let be the square of the multiple correlation coefficient between the k-th independent variable and the other independent variables. Let be the variance inflation factor of the k-th independent variable; where the variance inflation factor is used to measure the extent to which a single independent variable is explained by other independent variables. The larger the variance inflation factor, the more severe the collinearity.

[0069] Determine if the variance inflation factor exceeds a preset collinearity threshold. If it does, remove the independent variable; otherwise, retain the independent variable and output the candidate set of principal causes. For example... , indicating no collinearity. This indicates mild collinearity. If severe multicollinearity is indicated, the variable needs to be removed; Therefore, based on the above regression validation unit's assessment, the output conclusions include core cause robustness conclusions, interference variable exclusion conclusions, synergistic effect conclusions, and model reliability conclusions; for example, taking the aforementioned conclusions related to installation torque, temperature, and speed as an example... The robustness conclusion of the core cause: The regression coefficient of the installation torque is negative and there is no collinearity interference, indicating that the causal effect of "insufficient torque causing sensor abnormality" is robust and real, which is completely consistent with the PSM+ATE test results, and is identified as the main cause. Conclusion on exclusion of interfering variables: The regression coefficients of ambient temperature and collision speed are not significant, indicating that they have no real impact on sensor anomalies and can be excluded as possible causes. Conclusion on synergistic effects: The coefficients of the torque-temperature and torque-speed interaction terms are not significant, indicating that the effect of torque on anomalies is independent and there is no need to consider the synergistic effect of environment or speed. Model reliability conclusion: The model has no multicollinearity, the residuals follow a normal distribution, the fitting results are reliable, and it can be used as the core statistical basis for fault location.

[0070] The sensitivity analysis unit is used to quantify the relative contribution of high-potential variables to the uniform anomaly score, and outputs the main or compound causal factors as the causes of the failure; specifically: The coefficients of each independent variable in the multiple linear regression model are converted into standardized regression coefficients, and the relative contribution of each independent variable to the uniform anomaly score is output, expressed as follows:

[0071] in, The standardized coefficient is... Let the standard deviation of the independent variable be . To standardize the standard deviation of outlier scores, The coefficients of the independent variables; the relative contribution is calculated by dividing the absolute value of the standardized coefficient of a single independent variable by the sum of the absolute values ​​of the standardized coefficients of all independent variables; The SHAP value quantifies the local contribution of each independent variable to a single sample and its global contribution to all samples. The expression for the local contribution is:

[0072] in, For the multiple linear regression model for the first The actual predicted anomaly score for each sample To remove the current variable The post-multiple linear regression model's prediction anomaly score for the i-th sample; This refers to the SHAP value of a single sample. The global contribution expression is:

[0073] in, The SHAP values ​​for the entire sample. The number of samples; The magnitude of a variable’s influence on anomalies is determined based on single-sample SHAP values ​​and full-sample SHAP values, and the cause of the failure is output.

[0074] Therefore, based on the above sensitivity analysis process, the output conclusions include single-sample conclusions, global conclusions, and consistency conclusions. The single-sample conclusions output the main contributing causes of the fault based on the single-sample SHAP value. The global conclusions judge the global SHAP values ​​of all matched samples to verify whether the variable is the global main cause of sensor anomalies. The consistency conclusions are based on the complete consistency of the aforementioned prior screening, nearest neighbor matching, and regression verification conclusions, forming a quadruple verification.

[0075] The fault code generation module matches a preset fault code system based on the fault cause, hardware failure result, and group anomaly result, and outputs the fault code, cause description, and handling suggestions. For example, based on the aforementioned example, fault code 4021 is output, indicating an abnormal torque in the installation category - pelvic sensor. The handling suggestion is: reinstall the sensor to ensure the torque is within acceptable limits. Perform a functional check to verify the sensor response; if the abnormality persists, consider replacing the sensor and inspecting the mounting base.

[0076] like Figure 2 As shown, in another embodiment of this example, a method for intelligent diagnosis of dummy sensor anomalies is further included, applied to the aforementioned intelligent diagnosis system for dummy sensor anomalies, comprising: S1: After establishing a unique identification index for each dummy and its corresponding sensor, collect environmental test data and sensor data of the dummy and the sensor. S2: After preprocessing the sensor data, the impact segment data is truncated according to a fixed window length, and key features are extracted from the impact segment data to generate a feature vector; at the same time, the environmental test data is normalized. S3: Based on the rule layer, hardware faults are determined in the normalized environmental test data; based on the statistical layer, anomaly scores are calculated for sensor data anomalies; based on the model layer, anomaly scores are calculated for feature vectors to generate hardware fault results, statistical anomaly scores, and model anomaly scores. S4: Compare historical data from the same sensor and the same sensor on the same model of dummy to generate individual similarity scores and group anomaly results respectively; S5: Integrate statistical anomaly scores, model anomaly scores, and individual similarity scores to calculate a unified anomaly score; S6: Receives unified anomaly scores, generates target response quantities, extracts key variables to be verified, performs prior screening, nearest neighbor matching, regression verification and sensitivity analysis, and outputs fault causes; S7: Match the preset fault code system based on the fault cause, hardware fault result and group anomaly result, and output the fault code, cause description and handling suggestions.

[0077] Example 2: The difference between Embodiment 2 and Embodiment 1 is that Embodiment 2 also includes a sensor calibration scheduling module. In this module, the base cycle of abnormal sensor calibration scheduling is dynamically adjusted based on the usage frequency factor, environmental factor, and performance trend factor. The module also calls the intelligent scheduling model to generate calibration scheduling tasks based on the adjusted base cycle, abnormal sensor parameters, optimal target, and constraints.

[0078] In this embodiment, the basic cycle of the calibration plan is first adjusted. The basic cycle is set at 12 months based on industry standards. The adjustment process is affected by frequency factors, environmental factors, and performance trend factors. The frequency factor includes the usage frequency, which is divided into high frequency and low frequency. For example, high frequency means more than 8 tests per month, and the adjustment is to reduce the basic cycle by 2 months. Low frequency means less than 3 tests per month, and the adjustment is to extend it by 1 month. The environmental factor is environmental compensation. For example, if the device is used in a high temperature and high humidity environment for a long time, the basic cycle is reduced by 1 month. If the device is used in a normal environment, the basic cycle is not adjusted. The performance trend factor is the performance trend. For example, if the calibration deviation has been increasing in the last two calibrations, the basic cycle is reduced by 1-2 months. Conversely, if the calibration deviation is stable, the basic cycle is extended by 1 month.

[0079] Subsequently, an intelligent scheduling optimization model was constructed. The variables and parameters of this model include sensor parameters, resource parameters, and decision variables. For example, sensor parameters include the sensor set, sensor expiration month, sensor calibration time, risk weight, and allowable adjustment window. Resource parameters include regular working hours and overtime hours. The decision variable is whether the sensor is scheduled for calibration in a certain month.

[0080] The objective function of the model is to minimize the total overtime cost, expressed as:

[0081] in, The unit's overtime cost, for example, 50 yuan / hour. For overtime hours, For the m-th month; the core of the objective function is to make the most of regular working hours and reduce overtime expenses; The constraints for model construction include: Window constraints: Each sensor can only be calibrated within the allowed adjustment window to avoid excessively early or late calibration; Working time constraints: The total monthly calibration time shall not exceed the sum of regular working hours and overtime hours to avoid resource overload; High-risk priority constraint: High-risk sensors with a risk weight exceeding the threshold must be calibrated before the due month and cannot be delayed; Based on the above intelligent scheduling optimization model, for small-scale scenarios, such as fewer than 50 sensors, mixed-integer linear programming is used to obtain the global optimal solution; for large-scale scenarios, a priority-due-date greedy algorithm is used, with the following steps: 1. Sort the sensors in descending order of risk weight, giving priority to high-risk sensors; 2. Sort the sensors with the same priority in ascending order of due date; 3. Assign the sensors to months with sufficient working hours within the allowed window to quickly obtain an approximate optimal solution.

[0082] Example as follows: Assuming 20 sensors need to be calibrated in August (total time required: 40 people) (hourly), regular working hours: 100 people Hours, no overtime required; all 5 high-risk sensors are scheduled for completion in August; of the 15 low-risk sensors, 3 were rescheduled to July (to balance the remaining work hours in July), and 12 were scheduled for August. The total time spent in August = 5 × 2 + 12 × 2 = 34 person-hours. Hour people Hours, no overtime costs, meets all constraints.

[0083] After processing by the above-mentioned intelligent scheduling optimization model, the optimal calibration scheduling task is output. The system generates a standard calibration task sheet, which includes sensor ID, installation location, calibration requirements, and completion time limit. This sheet is pushed to maintenance personnel via mobile device. After arriving at the site, maintenance personnel scan the sensor ID to confirm the start of the task. After calibration is completed, the calibration deviation value and pass / fail status are entered and automatically linked to the equipment file.

[0084] 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 dummy sensor anomaly intelligent diagnostic system, characterized in that: It includes a data acquisition module, a preprocessing module, an anomaly detection module, an intelligent comparison module, an anomaly scoring module, a dynamic variable analysis module, and a fault code generation module, among which: The data acquisition module is used to collect environmental test data and sensor data of the dummies and sensors after establishing a unique identification index for each dummy and its corresponding sensor; The preprocessing module is used to preprocess the sensor data, extract the impact segment data according to a fixed window length, extract key features from the impact segment data, and generate feature vectors; the preprocessing module also performs normalization processing on the environmental test data; The anomaly detection module determines hardware faults in normalized environmental test data based on the rule layer, calculates anomaly scores for sensor data anomalies based on the statistics layer, and calculates anomaly scores for feature vectors based on the model layer, generating hardware fault results, statistical anomaly scores, and model anomaly scores. The intelligent comparison module is used to compare historical data of the same sensor and the same sensor position of the same type of dummy, and generate individual similarity scores and group anomaly results respectively. The anomaly scoring module is used to integrate statistical anomaly scores, model anomaly scores, and individual similarity scores to calculate a unified anomaly score; The dynamic variable analysis module is used to receive unified anomaly scores, generate target response quantities, extract key variables to be verified, perform prior screening, nearest neighbor matching, regression verification and sensitivity analysis, and output the causes of failures. The fault code generation module is used to match a preset fault code system based on the fault cause, hardware fault result, and group anomaly result, and output the fault code, cause description, and handling suggestions.

2. The intelligent diagnostic system for dummy sensor anomalies according to claim 1, characterized in that: The data acquisition module establishes a unique identity index for each dummy and its corresponding sensor as follows: Each dummy is assigned a unique ID. Sensors are numbered according to dummy ID, sensor model, and installation location, and basic sensor information is registered. The system collects and stores environmental test data, sensor data, and historical data from the crash test dummy. It also establishes a cross-data type association index using dummy ID, sensor number, and test time as core key fields to generate a multi-dimensional data association map.

3. The intelligent diagnostic system for dummy sensor anomalies according to claim 1, characterized in that: The anomaly detection module includes a rule layer unit, a statistical layer unit, and a model layer unit, wherein: The rule layer unit is used to receive preprocessed sensor data and directly determine hardware-level anomalies when there are constant values, NaN values, or over-range conditions, and output hardware fault results. The statistical layer unit calculates the standardized score of the feature vector based on the historical data distribution of sensors of the same model and location, and transforms it into anomaly scores. The expression is: in, These are abnormal scores. For the first The average of the standardized scores of the eigenvectors, For the first The threshold of the standardized scores of the feature vectors. For adjustment coefficients; For the first Standardized scores of each feature vector. For the first The values ​​of each eigenvector, The first in historical normal data The mean of the eigenvectors, The first in historical normal data The standard deviation of each eigenvector; The statistical layer unit also tracks the drift trend of the feature vector based on an exponentially weighted moving average and calculates the trend anomaly score of the feature vector. ; and based on the abnormal score and trend anomaly score Perform weighted summation to generate statistical outlier scores. The expression is: in, For trend anomaly scores, This indicates a cutoff function to prevent over-range drift from rendering the fractions meaningless; For the first The exponentially weighted moving average of the sensor target feature vector in this experiment. For the first The mean of historical normal data for each feature vector; These are the weighting coefficients. For the first The feature vector values ​​of this trial For the first The exponentially weighted moving average of the trials; Weights for outlier scores. Weights for trend anomaly scores; The model layer unit is used to train the autoencoder model based on historical sensor data. It inputs the feature vector into the autoencoder model, outputs a reconstructed vector, and calculates the reconstruction error, expressed as: Normalization to : in, For reconstruction error, The number of feature vectors, These are the original values ​​of the feature vector. The reconstructed values ​​of the feature vector, For model anomaly scores, This represents the truncation function. This represents the minimum reconstruction error of the feature vector within historical normal data. This represents the maximum reconstruction error of the feature vector in historical normal data.

4. The intelligent diagnostic system for dummy sensor anomalies according to claim 3, characterized in that: The intelligent comparison module includes an individual comparison unit and a group comparison unit, wherein: The individual comparison unit is used to compare historical data from the same sensor, calculate the DTW distance between the current waveform and the most recent N normal waveforms, and convert it into a similarity score. ; The group comparison unit is used to compare the feature vectors of sensors at the same position on the same type of dummy and calculate the cosine similarity. If the cosine similarity is lower than the preset threshold, it is marked as an abnormal individual in the group. If more than M sensors have a cosine similarity lower than the preset threshold, it is marked as a system abnormality and a group abnormality result is generated.

5. The intelligent diagnostic system for dummy sensor anomalies according to claim 4, characterized in that: The dynamic variable analysis module includes a variable definition unit, a priori screening unit, a nearest neighbor matching unit, a regression validation unit, and a sensitivity analysis unit, wherein: The variable definition unit is used to select a uniform anomaly score as an indicator to measure the degree of anomaly, generate the target response quantity, and extract candidate anomaly variables from environmental test data. The prior screening unit is used to calculate the correlation between the abnormal candidate variables and the target response, screen out the TOP-K variables, and output a set of high-potential variables. The nearest neighbor matching unit is used to perform propensity score matching calculations for individual high-potential variables in the set of high-potential variables as a treatment group, and to set up a control group based on the propensity score of the treatment group. The average treatment effect of the treatment group and the control group is calculated, and the potential principal causes of the abnormality are output. The regression validation unit uses the uniform anomaly score as the dependent variable and the high-potential variable as the independent variable to construct a multiple linear regression model to calculate the variance inflation factor of each independent variable and output robust principal causal candidates for causal effects. The sensitivity analysis unit is used to quantify the relative contribution of high-potential variables to the unified anomaly score and output the main or compound causal factors as the causes of failure.

6. The intelligent diagnostic system for dummy sensor anomalies according to claim 5, characterized in that: In the nearest neighbor matching unit, a propensity score matching calculation is performed on a single high-potential variable in the set of high-potential variables to serve as the treatment group. A control group is set up based on the propensity score of the treatment group. The average treatment effect of the treatment group and the control group is calculated, and the potential main causes of the anomaly are output as follows: Construct a logistic regression model with the single high-potential variable (whether the user belongs to the treatment group) as the dependent variable and other high-potential variables as independent variables. Calculate the propensity score for each dependent variable, expressed as follows: in, For the first The propensity score of each dependent variable. For conditional probability, when the dependent variable When fixed, the probability of belonging to the treatment group is... This is the intercept term for logistic regression. Represents the regression coefficient of the dependent variable. Indicates the first p dependent variable values ​​for a sample; Treatment groups are generated based on propensity scores. For each high-potential variable in the treatment group, a corresponding control group is matched, where the control group has the smallest difference in propensity score compared to the treatment group. Based on the matched samples, the average treatment effect between the treatment group and the control group was calculated. The expression is: in, This represents the number of samples in the processing group after nearest neighbor matching. This represents the number of samples in the control group after nearest neighbor matching; This represents the outlier response values ​​of all processed group samples after matching. Sum of abnormal response values To standardize anomaly scoring; Abnormal response values ​​of all control group samples after matching sum; Calculate the average treatment effect Standard error ,pass Test the calculation of average treatment effect of The value, expressed as: according to Value lookup The distribution table determines the degrees of freedom, and the values ​​can be looked up based on the degrees of freedom. Value, comparison Is the value less than the significance level? If so, this indicates that the sample is a potential primary cause of the anomaly.

7. The intelligent diagnostic system for dummy sensor anomalies according to claim 6, characterized in that: The regression validation unit uses a uniform anomaly score as the dependent variable and high-potential variables as independent variables to construct a multiple linear regression model, calculate the variance inflation factor of each independent variable, and output robust principal causal candidates for causal effects as follows: To standardize anomaly scoring Using the variable with high potential as the dependent variable and the variable with high potential as the independent variable, a multiple linear regression model is constructed, with the expression as follows: in, To standardize anomaly scoring, For the intercept term, The regression coefficients of the current independent variable are... For the current independent variable, For the first The regression coefficients of the independent variables, For the first One independent variable, This is the error term; Calculate the variance inflation factor between the individual independent variable and other independent variables. The expression is: in, Let be the square of the multiple correlation coefficient between the k-th independent variable and the other independent variables. Let be the variance inflation factor of the k-th independent variable; Determine whether the variance inflation factor exceeds the preset collinearity threshold. If it does, remove the independent variable; otherwise, retain the independent variable and output the candidate set of main causes.

8. The intelligent diagnostic system for dummy sensor anomalies according to claim 7, characterized in that: In the sensitivity analysis unit, the relative contribution of high-potential variables to the unified anomaly score is quantified, and the main or compound causal factors are output as the causes of the failure. The coefficients of each independent variable in the multiple linear regression model are converted into standardized regression coefficients, and the relative contribution of each independent variable to the uniform anomaly score is output, expressed as follows: in, The standardized coefficient is... Let the standard deviation of the independent variable be . To standardize the standard deviation of outlier scores, The coefficients of the independent variables; the relative contribution is calculated by dividing the absolute value of the standardized coefficient of a single independent variable by the sum of the absolute values ​​of the standardized coefficients of all independent variables; The SHAP value quantifies the local contribution of each independent variable to a single sample and its global contribution to all samples. The expression for the local contribution is: in, For the multiple linear regression model for the first The actual predicted anomaly score for each sample To remove the current variable The post-multiple linear regression model's prediction anomaly score for the i-th sample; This refers to the SHAP value of a single sample. The global contribution expression is: in, The SHAP values ​​for the entire sample. The number of samples; The magnitude of a variable’s influence on anomalies is determined based on single-sample SHAP values ​​and full-sample SHAP values, and the cause of the failure is output.

9. The intelligent diagnostic system for dummy sensor anomalies according to claim 8, characterized in that: It also includes a sensor calibration scheduling module, which dynamically adjusts the base cycle of abnormal sensor calibration scheduling based on usage frequency factors, environmental factors, and performance trend factors, and calls an intelligent scheduling model to generate calibration scheduling tasks based on the adjusted base cycle, abnormal sensor parameters, optimal targets, and constraints.

10. A method for intelligent diagnosis of dummy sensor anomalies, applied to the intelligent diagnosis system for dummy sensor anomalies as described in any one of claims 1-9, characterized in that: include: S1: After establishing a unique identification index for each dummy and its corresponding sensor, collect environmental test data and sensor data of the dummy and the sensor. S2: After preprocessing the sensor data, the impact segment data is truncated according to a fixed window length, and key features are extracted from the impact segment data to generate a feature vector; at the same time, the environmental test data is normalized. S3: Based on the rule layer, hardware faults are determined in the normalized environmental test data; based on the statistical layer, anomaly scores are calculated for sensor data anomalies; based on the model layer, anomaly scores are calculated for feature vectors to generate hardware fault results, statistical anomaly scores, and model anomaly scores. S4: Compare historical data from the same sensor and the same sensor on the same model of dummy to generate individual similarity scores and group anomaly results respectively; S5: Integrate statistical anomaly scores, model anomaly scores, and individual similarity scores to calculate a unified anomaly score; S6: Receives unified anomaly scores, generates target response quantities, extracts key variables to be verified, performs prior screening, nearest neighbor matching, regression verification and sensitivity analysis, and outputs fault causes; S7: Match the preset fault code system based on the fault cause, hardware fault result and group anomaly result, and output the fault code, cause description and handling suggestions.

Images