Product quality safety risk monitoring method based on inspection and detection data

Abstract

The invention discloses a product quality safety risk monitoring method based on inspection and detection data, and the method comprises the steps: collecting the inspection and detection data of an industrial part, carrying out the time sequence decomposition of the inspection and detection data, obtaining a qualified process parameter and a drifting process parameter, obtaining the local marginal influence of the process parameters on a quality index through nonlinear fitting, and carrying out the monitoring of the product quality safety risk. Performing cross-process correlation on the local marginal influence and the time series data set through a tense convolutional network to obtain global correlation characteristics, evolving continuous degradation based on the global correlation characteristics, the environmental parameters and the material attenuation characteristics of the industrial parts to obtain a part quality prediction result, and monitoring quality index fluctuation according to the quality prediction result and the process parameters to obtain a part quality prediction result. And obtaining part stability judgment and risk early warning levels. According to the method, global correlation characteristics are captured by using interference of drift process parameters on qualified working conditions, the precision and real-time performance of product quality safety risk monitoring are improved, and meanwhile, the method has relatively good interpretability.

Description

Technical Field

[0001] This invention relates to the field of risk monitoring, and in particular to a method for monitoring product quality and safety risks based on inspection and testing data. Background Technology

[0002] In industrial parts manufacturing, product quality and safety risk monitoring is a core link in ensuring the reliability of parts throughout their entire life cycle. With the popularization of inspection and testing technologies, the massive amount of time-series data accumulated during the production process provides a foundation for mining material degradation patterns and providing early warnings of failure risks. In recent years, machine learning has been gradually applied to this field, modeling quality process relationships through data-driven methods to achieve automated monitoring of product quality and safety risks.

[0003] Traditional quality monitoring methods, such as Shewhart control charts, rely heavily on manual experience or static threshold judgments, resulting in low accuracy and significant lag. While these methods can detect obvious quality anomalies, their ability to identify subtle changes such as process drift is limited. They cannot effectively meet the high-precision monitoring needs of the complex working conditions in modern automobile manufacturing. Furthermore, they are poorly adaptable to the dynamic and complex nature of data, making it difficult to quantify risks and provide early warnings. This, to some extent, restricts the development of product quality monitoring technology. Therefore, developing a product quality and safety risk monitoring method based on inspection and testing data has become an important way to improve the effectiveness of quality monitoring, quantify risks, and provide early warnings of safety anomalies. Summary of the Invention

[0004] The purpose of this invention is to provide a product quality and safety risk monitoring method based on inspection and testing data.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention provides a product quality and safety risk monitoring method based on inspection and testing data, comprising: The inspection and testing data of industrial parts are collected and preprocessed to obtain a time-series dataset containing process parameters, quality indicators and environmental parameters; Based on the time-series dataset, time-series decomposition is performed, and process parameters are divided according to the processing frequency fluctuations and temperature drifts obtained from the decomposition; qualified process parameters and drifting process parameters are obtained. The discrete data in the qualified process parameters are encoded into continuous values, and the continuous data are nonlinearly fitted using the Jacobian matrix. The local marginal impact of the process parameters on the quality indicators is obtained based on the fitting results. Local marginal effects and time-series datasets are correlated across processes using a temporal convolutional network. The drift process parameters are used to interfere with local marginal effects, and the network hole rate is adjusted based on the simulation error of the interference to obtain global correlation features. Based on the global correlation features, environmental parameters, and material degradation characteristics of industrial parts, a family of continuous degradation curves is generated. The probability distribution of the family of continuous degradation curves is statistically analyzed in terms of time dimension. Based on the probability distribution of material degradation, the part quality prediction result is obtained. Based on the quality prediction results and process parameters, monitor the fluctuations of quality indicators to obtain the stability judgment and risk warning level of the parts.

[0006] Furthermore, the method for obtaining the qualified process parameters and the drift process parameters includes: Based on the time series dataset, baseline data and material critical deviations corresponding to process parameters, quality indicators, and environmental parameters are obtained. Seasonal trend decomposition is performed on the time series dataset to obtain long-term trend terms. High-frequency detail coefficients are extracted from the time series dataset through wavelet transform. The root mean square of frequency fluctuation is calculated based on the high-frequency detail coefficients through a fixed-length window. If the window fluctuation amplitude exceeds 3% of the process baseline, it is marked as a disturbance window. Non-disturbance windows are divided into qualified process parameters. The time-varying fluctuation index is calculated using the difference ratio method based on the root mean square of the frequency fluctuation and the root mean square of the reference frequency. The cumulative drift deviation of the long-term trend term and the baseline data is calculated based on the disturbance window. If the time-varying fluctuation index within the disturbance window is less than or equal to 5% and the cumulative drift deviation is less than or equal to the material critical deviation, the disturbance window is added to the qualified process parameter; otherwise, it is classified as a drift process parameter.

[0007] Furthermore, the method for addressing local marginal effects includes: Based on the qualified process parameters, discrete data, continuous data, and corresponding qualified quality indicators are extracted. The discrete data is mapped to the continuous space through Walsh function orthogonal encoding. If the mutual information between the continuous vector and the discrete data after discrete encoding is greater than or equal to 0.85, the continuous vector is concatenated with the normalized continuous data vector according to the dimension. If the mutual information is less than 0.85, the continuous vector is weighted by 1.5 times and concatenated with the effective bits of the one-hot encoding. If the mutual information is still less than 0.6 after weighting, the learning rate of the continuous vector is set to 1 / 2 of the global learning rate when training the feedforward neural network. The discrete data includes the status of processing equipment and quality inspection classification labels. The concatenated continuous data vector is input into the input layer of the feedforward neural network. The qualified quality index is used as the qualified label. The hidden layer fits the nonlinear response through the hyperbolic tangent activation function. The feedforward neural network minimizes the residual through the backpropagation adaptive learning rate algorithm. If the root mean square error in the residual is less than or equal to 5%, the training is completed, and the working condition surface model and target quality index are obtained. The Jacobian matrix is ​​calculated based on the input layer samples and the target quality index. Based on the Jacobian matrix, the local high-order nonlinear effects of the samples are compensated through third-order Taylor expansion to obtain the stable working condition matrix. The matrix elements are used as local marginal effects. The basic disturbance condition matrix is ​​obtained by training the condition surface model based on the drift process parameters.

[0008] Furthermore, the method of using the drift process parameters to interfere with local marginal effects includes: Based on the working condition surface model, the input layer to hidden layer weight matrix and the hidden layer to output layer weight matrix are used as the initial drift coupling weight matrix. According to the initial drift coupling weight matrix, 30% of the working condition surface model weights are trained through the drift loss function to obtain the input side drift coupling weight matrix and the output side drift coupling weight matrix. The drift loss function is: ; in Here, MSE is the loss function, and the mean squared error is... To optimize the disturbance case matrix, the predicted values ​​are obtained from the basic disturbance case matrix. Based on the basic disturbance condition matrix, The regularization coefficient is . , The drift coupling weight matrix on the input side. This is the baseline weight matrix from the input layer to the hidden layer; The drift interference matrix is ​​obtained from the drift coupling weight matrix, the stable operating condition matrix, and the basic disturbance operating condition matrix using the drift interference matrix formula, which is: ; in For the drift interference matrix, To stabilize the operating condition matrix, It is the identity matrix. For drift process parameters, The mean of the baseline parameters, For Kronecker product, This is the transpose of the output-side drift coupling weight matrix; Cross-process parameters are extracted based on time-series datasets. The drift interference matrix and cross-process parameters are expanded into three-dimensional tensors and concatenated along the parameter dimensions to form an input tensor. The input tensor is then fed into a temporal convolutional network with three layers of dilated convolutions. The dilation rate is adjusted based on the simulated error of drift interference to capture cross-process dependencies layer by layer. The convolution results with different dilation rates are concatenated to obtain global correlation features and attention weight matrices.

[0009] Furthermore, the method for adjusting the network hole rate based on the simulation error includes: The target dilation rate is obtained based on the drift interference matrix, the stable condition matrix, and the attention weight matrix using a dilation rate adjustment formula. This target dilation rate is then used to adjust the receptive field of the next convolution in the temporal convolutional network. The dilation rate adjustment formula is as follows: ; in The target void ratio at time step t+1. Here, represents the hole rate at time step t, where t is the acquisition time step of the time series dataset, and Clip is the clipping function. This is the lower limit of the void ratio. This is the upper limit of the void ratio. The learning rate hyperparameter, , For the drift interference matrix, To stabilize the operating condition matrix, Here is the attention weight matrix. Global correlation features The 1D convolution is used, where softmax is the normalization function.

[0010] Furthermore, the method for obtaining the family of continuous degradation curves includes: Based on quality indicators, the material degradation characteristics of industrial parts are obtained. Then, based on these material degradation characteristics, global correlation features, and environmental parameters, a neural operator is used to calculate the time-varying rate of material degradation. The neural operator expression is as follows: ; in For neural operators, The weight of the hidden layer in the working condition surface model is 30%. The current state of degradation. For global correlation features, Here, M represents the environmental parameter, and M represents the material degradation characteristic. This is element-wise multiplication; The L2 normalized result of the stable operating condition matrix is ​​used as the initial degenerate state of the neural operator. Matrix multiplication and softmax normalization are performed on the drift interference matrix and the degenerate state to obtain time-varying modulation coefficients. Time-varying hyperbolic tangent functions are generated using the tanh function based on the time-varying modulation coefficients. The hidden activation function of the operating condition surface model is replaced with the time-varying hyperbolic tangent function. A family of continuous degenerate curves within the prediction time is generated based on the time-varying operating condition surface model using the adaptive step-size Euler method.

[0011] Furthermore, the method for obtaining the predicted quality result of the part includes: Based on the family of continuous degradation curves, the prediction time is divided into time windows. The degradation parameters within the time window are sequentially concatenated into a state vector. The predicted value of the quality index is obtained through the working condition surface model based on the state vector. The probability distribution of the degradation state is estimated by Gaussian kernel density based on the state vector. The drift disturbance matrix is ​​converted into a diagonal matrix. The element-wise multiplication of the stable condition matrix and the diagonal matrix is ​​calculated. The element-wise multiplication and the L1 norm of the state vector are used as the instantaneous failure rate. Based on the probability distribution, the part quality prediction result is obtained through a material degradation risk accumulation algorithm. The part quality prediction result includes cumulative failure risk and predicted quality index values. The algorithm expression is as follows: ; in To represent the cumulative failure risk, Y is the prediction duration. This is the region of failure. The set of numbers less than or equal to Q for At any moment, in a state of degradation The instantaneous failure rate, For the current time window, for The historical moment before that moment, To obtain through probability distribution At any moment, in a state of degradation The probability density under, Let Q be the predicted value of the quality indicator, and let Q be the failure threshold of the quality indicator. for standard deviation The standard normal cumulative distribution function is calculated using the ERF error function.

[0012] Furthermore, the method for obtaining the stability determination of the component and the risk warning level includes: Based on the qualified process parameters, obtain the corresponding historical data of quality indicators, calculate the mean and standard deviation of the historical data of quality indicators, use the mean as the center line, determine the upper control limit and lower control limit according to the three-times-standard-deviation principle, remove outliers that exceed the control limit, and recalculate the mean and standard deviation until 99.73% of the data fall within the control limit, and obtain the Shewhart control chart model based on the control limit. Input the predicted values ​​of the quality indicators into the Shewhart control chart model. If the predicted values ​​of the quality indicators exceed the control limits, the model is considered unstable. If the time-varying fluctuation index of the process parameters is greater than 5%, the model is considered unstable. Based on the cumulative failure risk, green, yellow, and red warnings are output according to the industry risk grading standard range. If the quality result is unstable, the warning level will be raised by one level.

[0013] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: This invention obtains the local marginal impact of process parameters on quality indicators through nonlinear fitting, explores the correlation characteristics between process parameters and quality indicators, reflects the quality change law in complex production processes, and enhances the model's ability to fuse and process multi-source heterogeneous data by dynamically adjusting the network hole rate based on the simulation error of drift interference through a temporal convolutional network. This improves the sensitivity and capture ability of quality fluctuations. Based on global correlation features, environmental parameters, and material attenuation characteristics, it derives a family of continuous degradation curves for parts and uses this to predict part quality, improving the real-time performance of quality risk warnings. Combined with the Shewhart control chart model, it determines the stability of the predicted quality indicator values ​​and classifies risk warning levels, improving the interpretability of risk assessment and making the risk warning results more credible and practical. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the steps of a product quality and safety risk monitoring method based on inspection and testing data, as described in an embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0016] Reference Figure 1 As shown, the present invention provides a product quality and safety risk monitoring method based on inspection and testing data, comprising: The inspection and testing data of industrial parts are collected and preprocessed to obtain a time-series dataset containing process parameters, quality indicators and environmental parameters; In the actual evaluation, a certain automobile engine cylinder head die-casting production line was selected as the object. 72 hours of continuous production data were collected. Die-casting temperature, hydraulic pump operating frequency, holding time, mold temperature, and cooling water flow rate were collected as process parameters. Flatness, wall thickness difference, hardness, processing equipment status, quality inspection classification labels, and material attenuation characteristics were collected as part quality indicators. Temperature, humidity, and air pressure were collected as environmental parameters. The process parameters and environmental parameters were collected once per minute. The quality indicators were detected offline at the batch level. One batch was produced every 30 minutes. The cylinder head was measured by a coordinate measuring machine and the quality value was recorded. The batch quality value was copied to all time points within that minute through production cycle mapping. Finally, a time series dataset of 4320 time steps was obtained. Based on the time-series dataset, time-series decomposition is performed, and process parameters are divided according to the processing frequency fluctuations and temperature drifts obtained from the decomposition; qualified process parameters and drifting process parameters are obtained. In the actual evaluation, STL decomposition was performed on the die-casting temperature based on the time-series dataset to separate the long-term trend term. From hour 30 to 40, the die-casting temperature slowly increased from 600℃ to 603℃, with a cumulative drift of 3℃, which is less than the critical deviation of 5℃. Wavelet decomposition was performed on the time-series data after removing the trend term to extract the high-frequency detail coefficients of the high-frequency oscillations of the hydraulic system of the fifth-layer die-casting machine. Based on the high-frequency detail coefficients, the root mean square (RMS) of the high-frequency oscillations within each 10-minute window was calculated using a sliding window of fixed length. The RMS of window 1 was 0.06Hz, which is greater than the 3% threshold of 0.05Hz above the process baseline, thus marking the perturbation window. The RMS of window 35 was 0.04Hz. The window is retained as a qualified process window. The root mean square of the frequency fluctuation of the disturbance window is subtracted from the root mean square of the reference frequency. Then, the difference is divided by the root mean square of the reference frequency. The percentage of the absolute value of the difference is calculated to obtain the time-varying fluctuation index of 20%, which is greater than the fluctuation threshold of 5%. The long-term trend of the mold temperature in the corresponding period of window 1 rises from 225℃ to 227℃, with a cumulative drift of 2℃, which is less than the critical deviation of 5℃. Window 1 is divided into drift process parameters. The discrete data in the qualified process parameters are encoded into continuous values, and the continuous data are nonlinearly fitted using the Jacobian matrix. The local marginal impact of the process parameters on the quality indicators is obtained based on the fitting results. In the actual evaluation, the three discrete values ​​(0,1,2) of the processing equipment status of the discrete data in the qualified process parameters are mapped to a 3-dimensional orthogonal continuous space through the Walsh function to generate an encoding vector. The mutual information between the encoding vector and the original discrete value is 0.92, which is greater than 0.85, satisfying the splicing condition. The continuous data such as die casting temperature and wall thickness difference are normalized by min-max. The continuous vector is spliced ​​with the normalized continuous data vector according to the dimension to obtain a 6-dimensional continuous data vector. The concatenated continuous data vector is input into the input layer of a feedforward neural network. The qualified quality indicators are used as the qualified labels. The structure of the feedforward neural network input layer is a continuous data vector dimension, with 16 neurons in the hidden layer. The nonlinear response is fitted using the hyperbolic tangent activation function tanh. The output layer outputs 3-dimensional data corresponding to the three quality indicators of the material. The Adam optimizer is used with an adaptive learning rate. The initial learning rate is set to 0.001. The mean square error is minimized using a loss function, with the objective being that the residual is less than 5% of the quality indicator. For example, the training objective for flatness is 0.0025 mm. After training, the working condition surface model and the target quality indicators are obtained, which can map the input vector to the quality indicators. The Jacobian matrix is ​​calculated for the input and output vectors using the autograd tool. A third-order Taylor expansion of the Jacobian matrix is ​​performed, and the second and third derivatives are calculated to obtain the compensation prediction values. The local higher-order nonlinear effects of the compensation samples are then used to obtain the stable operating condition matrix, whose elements represent the local marginal influence of process parameters on quality indicators. The die-casting temperature is... For every 1°C increase in die-casting temperature, the flatness increases by 0.0005 mm; The drift process parameters are input into the working condition surface model, the existing network structure is reused, only some parameters are fine-tuned, and 30% of the hidden layer weights are trained to obtain the basic perturbation working condition moment, which is used as the deviation of the quality index under the drift working condition, where the flatness deviation increases by 0.001mm / ℃. Local marginal effects and time-series datasets are correlated across processes using a temporal convolutional network. The drift process parameters are used to interfere with local marginal effects, and the network hole rate is adjusted based on the simulation error of the interference to obtain global correlation features. In the actual evaluation, based on the working condition surface model, the input layer to hidden layer weight matrix was used as the initial input-side drift coupling weight matrix, and the hidden layer to output layer weight matrix was used as the initial output-side drift coupling weight matrix. The initial learning rate was set to 0.001, and the regularization coefficient was set to 0.05. Only 30% of the parameters in the first 5 layers of the drift coupling weight matrix were updated. After 100 iterations, the mean squared error decreased from the initial 0.001 to... Weight bias Stabilize the weights within 0.01 to obtain the input-side drift coupling weight matrix and the output-side drift coupling weight matrix; The drift disturbance matrix is ​​obtained from the drift coupling weight matrix, the stable operating condition matrix, and the basic disturbance operating condition matrix using the drift disturbance matrix formula. The drift of five parameters, including die-casting temperature and holding time, is as follows: The norm of the benchmark parameter is , After ReLU activation to 0.02, the drift disturbance increases the local marginal impact by 10%, resulting in the final drift disturbance matrix. ; Based on the time-series dataset, cross-process parameters such as die-casting temperature, holding time, mold temperature, cooling water flow rate, and workshop temperature are extracted and expanded into a 4320x5 two-dimensional tensor. The drift interference matrix is ​​also expanded into a 4320x5 two-dimensional tensor. These are then concatenated along the parameter dimensions to obtain a 4320x10 three-dimensional tensor. This three-dimensional tensor is input into a TCN, and three layers of dilated convolutions are set, with an initial void ratio of... =1, It is 2. The target void ratio is 4, obtained through a void ratio adjustment formula based on the drift disturbance matrix, stable operating condition matrix, and attention weight matrix. It is 3. The target dilation rate is set to 5. The receptive field of the temporal convolutional network (TCN) for the next convolution is adjusted. After adjustment, the receptive field of the TCN expands from 15 to 25, more efficiently capturing cross-process drift correlations. Convolution results with different dilation rates are concatenated to obtain 192-dimensional multi-scale features. These features are projected onto a 64-dimensional space through linear transformation. The attention weight matrix B is obtained by scaling the dot product attention. Finally, the Value matrix of the multi-scale correlation features is fused with the time step weights to obtain global correlation features. It mainly integrates the local correlation during the drift period and the baseline correlation during the historical stable period to accurately characterize the time-varying correlation under drift interference; Based on the global correlation features, environmental parameters, and material degradation characteristics of industrial parts, a family of continuous degradation curves is generated. The probability distribution of the family of continuous degradation curves is statistically analyzed in terms of time dimension. Based on the probability distribution of material degradation, the part quality prediction result is obtained. In the actual evaluation, the working condition surface model contains three hidden layers, specifically 64, 32, and 16 neurons. When calculating the neural operator, five of the first 16 neurons are reused. Based on the drift interference matrix [1.2, 1.1, 1.05, 1.03, 1.02] and the degenerate state of the neural operator pair, matrix multiplication and softmax normalization are performed to obtain the time-varying modulation coefficients [0.3, 0.2, 0.2, 0.15, 0.15]. These time-varying modulation coefficients are then used to generate a time-varying hyperbolic tangent function using the tanh function. This time-varying hyperbolic tangent function replaces the hidden layer activation function of the working condition surface model. Based on the time-varying working condition surface model, a family of continuous degradation curves with a prediction time of 100 hours is generated using the adaptive step-size Euler method. The time step is 1 hour, and the prediction time covers the production time and 28 hours of aging simulation time. For a certain cylinder head made of aluminum alloy 6061, the hardness degradation curve under stable working conditions decreases by 0.01 HB per hour. Under drift conditions, the die-casting temperature is 615℃, and when the die-casting temperature drifts by 15℃, the hardness degradation curve decreases by 0.0325 HB per hour. Under stable working conditions, the flatness degradation curve increases by 0.00004 mm per hour, and under drift conditions, it increases by 0.00054 mm per hour. Based on a family of continuous degradation curves, the prediction time is divided into 10-hour time windows. The flatness increment and hardness decrease within each time window are sequentially concatenated into a state vector. Based on this state vector, the predicted quality index is obtained using a working condition surface model. The first 10 hours represent a stable working condition, with a predicted flatness quality value of 0.0504 mm. From the 20th to the 70th hour, a drift condition is observed, with a predicted value of 0.0704 mm, exceeding the failure threshold of 0.07 mm, indicating the approaching failure edge. The probability distribution of the state vector is estimated using a Gaussian kernel density estimation with 1000 Monte Carlo sampling points and a kernel bandwidth of 0.1, yielding a probability distribution of 15% for the current degradation state. The instantaneous failure rate under a certain state is calculated to be 2.463% / hour. Based on this probability distribution, a material degradation risk accumulation algorithm is used to obtain the part quality prediction result. Under the drift condition, the cumulative failure risk is 0.369% after 1 hour and 75% after 50 hours, while under the stable condition, it will only reach 75% after 500 hours. Based on the quality prediction results and process parameters, monitor the fluctuations of quality indicators to obtain the stability judgment and risk warning level of the parts.

[0017] In the actual evaluation, 500 stable working conditions of historical flatness data were extracted from the qualified process parameters. Among them, the cylinder head was produced continuously for 10 days, with 50 qualified batches per day. After removing the drift window, the mean of the historical quality index data was calculated to be 0.05mm and the standard deviation was 0.002mm. The mean was used as the center line. According to the principle of three times the standard deviation, the upper control limit was determined to be 0.056mm and the lower control limit was 0.044mm. Three points in the historical flatness data exceeded the UCL control limit and were removed and recalculated. The predicted values ​​of the quality indicators were input into the Shewhart control chart model. The predicted flatness value for 50 hours under drift conditions was 0.0704 mm, which exceeded the UCL and was judged as unstable in quality. The time-varying fluctuation index of the die casting temperature drift of 15℃ was 20%, which was greater than 5% and was judged as process drift and unstable. According to the industry risk classification standard, the risk value is green for [0,30%), yellow for [30%,60%), orange for [60%,80%), and red for [80%,100%). The cumulative failure risk of 75% is orange and is judged as unstable, upgraded to a red warning.

[0018] In this embodiment, the method for obtaining the qualified process parameters and the drift process parameters includes: Based on the time series dataset, baseline data and material critical deviations corresponding to process parameters, quality indicators, and environmental parameters are obtained. Seasonal trend decomposition is performed on the time series dataset to obtain long-term trend terms. High-frequency detail coefficients are extracted from the time series dataset through wavelet transform. The root mean square of frequency fluctuation is calculated based on the high-frequency detail coefficients through a fixed-length window. If the window fluctuation amplitude exceeds 3% of the process baseline, it is marked as a disturbance window. Non-disturbance windows are divided into qualified process parameters. The time-varying fluctuation index is calculated using the difference ratio method based on the root mean square of the frequency fluctuation and the root mean square of the reference frequency. The cumulative drift deviation of the long-term trend term and the baseline data is calculated based on the disturbance window. If the time-varying fluctuation index within the disturbance window is less than or equal to 5% and the cumulative drift deviation is less than or equal to the material critical deviation, the disturbance window is added to the qualified process parameter; otherwise, it is classified as a drift process parameter.

[0019] In this embodiment, the method for addressing local marginal effects includes: Based on the qualified process parameters, discrete data, continuous data, and corresponding qualified quality indicators are extracted. The discrete data is mapped to the continuous space through Walsh function orthogonal encoding. If the mutual information between the continuous vector and the discrete data after discrete encoding is greater than or equal to 0.85, the continuous vector is concatenated with the normalized continuous data vector according to the dimension. If the mutual information is less than 0.85, the continuous vector is weighted by 1.5 times and concatenated with the effective bits of the one-hot encoding. If the mutual information is still less than 0.6 after weighting, the learning rate of the continuous vector is set to 1 / 2 of the global learning rate when training the feedforward neural network. The discrete data includes the status of processing equipment and quality inspection classification labels. The concatenated continuous data vector is input into the input layer of the feedforward neural network. The qualified quality index is used as the qualified label. The hidden layer fits the nonlinear response through the hyperbolic tangent activation function. The feedforward neural network minimizes the residual through the backpropagation adaptive learning rate algorithm. If the root mean square error in the residual is less than or equal to 5%, the training is completed, and the working condition surface model and target quality index are obtained. The Jacobian matrix is ​​calculated based on the input layer samples and the target quality index. Based on the Jacobian matrix, the local high-order nonlinear effects of the samples are compensated through third-order Taylor expansion to obtain the stable working condition matrix. The matrix elements are used as local marginal effects. The basic disturbance condition matrix is ​​obtained by training the condition surface model based on the drift process parameters.

[0020] In this embodiment, the method of using the drift process parameters to interfere with local marginal effects includes: Based on the working condition surface model, the input layer to hidden layer weight matrix and the hidden layer to output layer weight matrix are used as the initial drift coupling weight matrix. According to the initial drift coupling weight matrix, 30% of the working condition surface model weights are trained through the drift loss function to obtain the input side drift coupling weight matrix and the output side drift coupling weight matrix. The drift loss function is: ; in Here, MSE is the loss function, and the mean squared error is... To optimize the disturbance case matrix, the predicted values ​​are obtained from the basic disturbance case matrix. Based on the basic disturbance condition matrix, The regularization coefficient is . , The drift coupling weight matrix on the input side. This is the baseline weight matrix from the input layer to the hidden layer; The drift interference matrix is ​​obtained from the drift coupling weight matrix, the stable operating condition matrix, and the basic disturbance operating condition matrix using the drift interference matrix formula, which is: ; in For the drift interference matrix, To stabilize the operating condition matrix, It is the identity matrix. For drift process parameters, The mean of the baseline parameters, For Kronecker product, This is the transpose of the output-side drift coupling weight matrix; Cross-process parameters are extracted based on time-series datasets. The drift interference matrix and cross-process parameters are expanded into three-dimensional tensors and concatenated along the parameter dimensions to form an input tensor. The input tensor is then fed into a temporal convolutional network with three layers of dilated convolutions. The dilation rate is adjusted based on the simulated error of drift interference to capture cross-process dependencies layer by layer. The convolution results with different dilation rates are concatenated to obtain global correlation features and attention weight matrices.

[0021] In this embodiment, the method for adjusting the network hole rate based on the simulation error includes: The target dilation rate is obtained based on the drift interference matrix, the stable condition matrix, and the attention weight matrix using a dilation rate adjustment formula. This target dilation rate is then used to adjust the receptive field of the next convolution in the temporal convolutional network. The dilation rate adjustment formula is as follows: ; in The target void ratio at time step t+1. Here, represents the hole rate at time step t, where t is the acquisition time step of the time series dataset, and Clip is the clipping function. This is the lower limit of the void ratio. This is the upper limit of the void ratio. The learning rate hyperparameter, , For the drift interference matrix, To stabilize the operating condition matrix, Here is the attention weight matrix. Global correlation features The 1D convolution is used, where softmax is the normalization function.

[0022] In this embodiment, the method for obtaining the family of continuous degradation curves includes: Based on quality indicators, the material degradation characteristics of industrial parts are obtained. Then, based on these material degradation characteristics, global correlation features, and environmental parameters, a neural operator is used to calculate the time-varying rate of material degradation. The neural operator expression is as follows: ; in For neural operators, The weight of the hidden layer in the working condition surface model is 30%. The current state of degradation. For global correlation features, Here, M represents the environmental parameter, and M represents the material degradation characteristic. This is element-wise multiplication; The L2 normalized result of the stable operating condition matrix is ​​used as the initial degenerate state of the neural operator. Matrix multiplication and softmax normalization are performed on the drift interference matrix and the degenerate state to obtain time-varying modulation coefficients. Time-varying hyperbolic tangent functions are generated using the tanh function based on the time-varying modulation coefficients. The hidden activation function of the operating condition surface model is replaced with the time-varying hyperbolic tangent function. A family of continuous degenerate curves within the prediction time is generated based on the time-varying operating condition surface model using the adaptive step-size Euler method.

[0023] In this embodiment, the method for obtaining the predicted quality result of the part includes: Based on the family of continuous degradation curves, the prediction time is divided into time windows. The degradation parameters within the time window are sequentially concatenated into a state vector. The predicted value of the quality index is obtained through the working condition surface model based on the state vector. The probability distribution of the degradation state is estimated by Gaussian kernel density based on the state vector. The drift disturbance matrix is ​​converted into a diagonal matrix. The element-wise multiplication of the stable condition matrix and the diagonal matrix is ​​calculated. The element-wise multiplication and the L1 norm of the state vector are used as the instantaneous failure rate. Based on the probability distribution, the part quality prediction result is obtained through a material degradation risk accumulation algorithm. The part quality prediction result includes cumulative failure risk and predicted quality index values. The algorithm expression is as follows: ; in To represent the cumulative failure risk, Y is the prediction duration. This is the region of failure. The set of numbers less than or equal to Q for At any moment, in a state of degradation The instantaneous failure rate, For the current time window, for The historical moment before that moment, To obtain through probability distribution At any moment, in a state of degradation The probability density under, Let Q be the predicted value of the quality indicator, and let Q be the failure threshold of the quality indicator. for standard deviation The standard normal cumulative distribution function is calculated using the ERF error function.

[0024] In this embodiment, the method for obtaining the stability determination of the component and the risk warning level includes: Based on the qualified process parameters, obtain the corresponding historical data of quality indicators, calculate the mean and standard deviation of the historical data of quality indicators, use the mean as the center line, determine the upper control limit and lower control limit according to the three-times-standard-deviation principle, remove outliers that exceed the control limit, and recalculate the mean and standard deviation until 99.73% of the data fall within the control limit, and obtain the Shewhart control chart model based on the control limit. Input the predicted values ​​of the quality indicators into the Shewhart control chart model. If the predicted values ​​of the quality indicators exceed the control limits, the model is considered unstable. If the time-varying fluctuation index of the process parameters is greater than 5%, the model is considered unstable. Based on the cumulative failure risk, green, yellow, and red warnings are output according to the industry risk grading standard range. If the quality result is unstable, the warning level will be raised by one level.

[0025] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A method for monitoring product quality and safety risks based on inspection and testing data, characterized in that, Includes the following steps: The inspection and testing data of industrial parts are collected and preprocessed to obtain a time-series dataset containing process parameters, quality indicators and environmental parameters; Based on the time-series dataset, time-series decomposition is performed, and process parameters are divided according to the processing frequency fluctuations and temperature drifts obtained from the decomposition; qualified process parameters and drifting process parameters are obtained. The discrete data in the qualified process parameters are encoded into continuous values, and the continuous data are nonlinearly fitted using the Jacobian matrix. The local marginal impact of the process parameters on the quality indicators is obtained based on the fitting results. Local marginal effects and time-series datasets are correlated across processes using a temporal convolutional network. The drift process parameters are used to interfere with local marginal effects, and the network hole rate is adjusted based on the simulation error of the interference to obtain global correlation features. Based on the global correlation features, environmental parameters, and material degradation characteristics of industrial parts, a family of continuous degradation curves is generated. The probability distribution of the family of continuous degradation curves is statistically analyzed in terms of time dimension. Based on the probability distribution of material degradation, the part quality prediction result is obtained. Based on the quality prediction results and process parameters, monitor the fluctuations of quality indicators to obtain the stability judgment and risk warning level of the parts.

2. The product quality and safety risk monitoring method based on inspection and testing data according to claim 1, characterized in that, The method for obtaining the qualified process parameters and the drift process parameters includes: Based on the time series dataset, baseline data and material critical deviations corresponding to process parameters, quality indicators, and environmental parameters are obtained. Seasonal trend decomposition is performed on the time series dataset to obtain long-term trend terms. High-frequency detail coefficients are extracted from the time series dataset through wavelet transform. The root mean square of frequency fluctuation is calculated based on the high-frequency detail coefficients through a fixed-length window. If the window fluctuation amplitude exceeds 3% of the process baseline, it is marked as a disturbance window. Non-disturbance windows are divided into qualified process parameters. The time-varying fluctuation index is calculated using the difference ratio method based on the root mean square of the frequency fluctuation and the root mean square of the reference frequency. The cumulative drift deviation of the long-term trend term and the baseline data is calculated based on the disturbance window. If the time-varying fluctuation index within the disturbance window is less than or equal to 5% and the cumulative drift deviation is less than or equal to the material critical deviation, the disturbance window is added to the qualified process parameter; otherwise, it is classified as a drift process parameter.

3. The product quality and safety risk monitoring method based on inspection and testing data according to claim 1, characterized in that, The method for addressing local marginal effects includes: Based on the qualified process parameters, discrete data, continuous data, and corresponding qualified quality indicators are extracted. The discrete data is mapped to the continuous space through Walsh function orthogonal encoding. If the mutual information between the continuous vector and the discrete data after discrete encoding is greater than or equal to 0.85, the continuous vector is concatenated with the normalized continuous data vector according to the dimension. If the mutual information is less than 0.85, the continuous vector is weighted by 1.5 times and concatenated with the effective bits of the one-hot encoding. If the mutual information is still less than 0.6 after weighting, the learning rate of the continuous vector is set to 1 / 2 of the global learning rate when training the feedforward neural network. The discrete data includes the status of processing equipment and quality inspection classification labels. The concatenated continuous data vector is input into the input layer of the feedforward neural network. The qualified quality index is used as the qualified label. The hidden layer fits the nonlinear response through the hyperbolic tangent activation function. The feedforward neural network minimizes the residual through the backpropagation adaptive learning rate algorithm. If the root mean square error in the residual is less than or equal to 5%, the training is completed, and the working condition surface model and target quality index are obtained. The Jacobian matrix is ​​calculated based on the input layer samples and the target quality index. Based on the Jacobian matrix, the local high-order nonlinear effects of the samples are compensated through third-order Taylor expansion to obtain the stable working condition matrix. The matrix elements are used as local marginal effects. The basic disturbance condition matrix is ​​obtained by training the condition surface model based on the drift process parameters.

4. The product quality and safety risk monitoring method based on inspection and testing data according to claim 1, characterized in that, The method of interfering with local marginal effects using the drift process parameters includes: Based on the working condition surface model, the input layer to hidden layer weight matrix and the hidden layer to output layer weight matrix are used as the initial drift coupling weight matrix. According to the initial drift coupling weight matrix, 30% of the working condition surface model weights are trained through the drift loss function to obtain the input side drift coupling weight matrix and the output side drift coupling weight matrix. The drift loss function is: ； in Here, MSE is the loss function, and the mean squared error is... To optimize the disturbance case matrix, the predicted values ​​are obtained from the basic disturbance case matrix. Based on the basic disturbance condition matrix, The regularization coefficient is . , The drift coupling weight matrix on the input side. This is the baseline weight matrix from the input layer to the hidden layer; The drift interference matrix is ​​obtained from the drift coupling weight matrix, the stable operating condition matrix, and the basic disturbance operating condition matrix using the drift interference matrix formula, which is: ； in For the drift interference matrix, To stabilize the operating condition matrix, It is the identity matrix. For drift process parameters, The mean of the baseline parameters, For Kronecker product, This is the transpose of the output-side drift coupling weight matrix; Cross-process parameters are extracted based on time-series datasets. The drift interference matrix and cross-process parameters are expanded into three-dimensional tensors and concatenated along the parameter dimensions to form an input tensor. The input tensor is then fed into a temporal convolutional network with three layers of dilated convolutions. The dilation rate is adjusted based on the simulated error of drift interference to capture cross-process dependencies layer by layer. The convolution results with different dilation rates are concatenated to obtain global correlation features and attention weight matrices.

5. The product quality and safety risk monitoring method based on inspection and testing data according to claim 4, characterized in that, The method for adjusting the network hole rate based on the simulation error includes: The target dilation rate is obtained based on the drift interference matrix, the stable condition matrix, and the attention weight matrix using a dilation rate adjustment formula. This target dilation rate is then used to adjust the receptive field of the next convolution in the temporal convolutional network. The dilation rate adjustment formula is as follows: ； in The target void ratio at time step t+1. Here, represents the hole rate at time step t, where t is the acquisition time step of the time series dataset, and Clip is the clipping function. This is the lower limit of the void ratio. This is the upper limit of the void ratio. The learning rate hyperparameter, , For the drift interference matrix, To stabilize the operating condition matrix, Here is the attention weight matrix. Global correlation features The 1D convolution is used, where softmax is the normalization function.

6. The product quality and safety risk monitoring method based on inspection and testing data according to claim 1, characterized in that, The method for obtaining the family of continuous degradation curves includes: Based on quality indicators, the material degradation characteristics of industrial parts are obtained. Then, based on these material degradation characteristics, global correlation features, and environmental parameters, a neural operator is used to calculate the time-varying rate of material degradation. The neural operator expression is as follows: ； in For neural operators, The weight of the hidden layer in the working condition surface model is 30%. The current state of degradation. For global association features, Here, M represents the environmental parameter, and M represents the material degradation characteristic. This is element-wise multiplication; The L2 normalized result of the stable operating condition matrix is ​​used as the initial degenerate state of the neural operator. Matrix multiplication and softmax normalization are performed on the drift interference matrix and the degenerate state to obtain time-varying modulation coefficients. Time-varying hyperbolic tangent functions are generated using the tanh function based on the time-varying modulation coefficients. The hidden activation function of the operating condition surface model is replaced with the time-varying hyperbolic tangent function. A family of continuous degenerate curves within the prediction time is generated based on the time-varying operating condition surface model using the adaptive step-size Euler method.

7. The product quality and safety risk monitoring method based on inspection and testing data according to claim 1, characterized in that, The method for obtaining the predicted quality result of the part includes: Based on the family of continuous degradation curves, the prediction time is divided into time windows. The degradation parameters within the time window are sequentially concatenated into a state vector. The predicted value of the quality index is obtained through the working condition surface model based on the state vector. The probability distribution of the degradation state is estimated by Gaussian kernel density based on the state vector. The drift disturbance matrix is ​​converted into a diagonal matrix. The element-wise multiplication of the stable condition matrix and the diagonal matrix is ​​calculated. The element-wise multiplication and the L1 norm of the state vector are used as the instantaneous failure rate. Based on the probability distribution, the part quality prediction result is obtained through a material degradation risk accumulation algorithm. The part quality prediction result includes cumulative failure risk and predicted quality index values. The algorithm expression is as follows: ； in To represent the cumulative failure risk, Y is the prediction duration. This is the region of failure. The set of numbers less than or equal to Q for At any moment, in a state of degradation The instantaneous failure rate, For the current time window, for The historical moment before that moment, To obtain through probability distribution At any moment, in a state of degradation The probability density under, Let Q be the predicted value of the quality indicator, and let Q be the failure threshold of the quality indicator. for standard deviation The standard normal cumulative distribution function is calculated using the ERF error function.

8. The product quality and safety risk monitoring method based on inspection and testing data according to claim 1, characterized in that, The method for obtaining the stability determination of the component and the risk warning level includes: Based on the qualified process parameters, obtain the corresponding historical data of quality indicators, calculate the mean and standard deviation of the historical data of quality indicators, use the mean as the center line, determine the upper control limit and lower control limit according to the three-times-standard-deviation principle, remove outliers that exceed the control limit, and recalculate the mean and standard deviation until 99.73% of the data fall within the control limit, and obtain the Shewhart control chart model based on the control limit. Input the predicted values ​​of the quality indicators into the Shewhart control chart model. If the predicted values ​​of the quality indicators exceed the control limits, the model is considered unstable. If the time-varying fluctuation index of the process parameters is greater than 5%, the model is considered unstable. Based on the cumulative failure risk, green, yellow, and red warnings are output according to the industry risk grading standard range. If the quality result is unstable, the warning level will be raised by one level.

Images