A vehicle-mounted camera cold forging production quality monitoring method based on multi-source data fusion

Abstract

The application relates to the technical field of production data processing, in particular to a vehicle-mounted camera cold forging production quality monitoring method based on multi-source data fusion, which comprises the following steps: collecting multi-source high-frequency data of a cold forging process and extracting deep physical features; based on historical ideal sample data, an ideal flow multi-sensor time sequence reference mode adaptive to working conditions is constructed; a multi-dimensional flow deviation of the deep physical features of a current stamping cycle and the reference mode is calculated in real time, and a statistical model trained based on historical defect data is used to calculate a flow deviation risk index; the internal defect risk level of a forged piece is evaluated according to the flow deviation risk index, and diagnostic information and closed-loop feedback are provided. Through the technical scheme, the internal defect risk of a cold forged piece can be evaluated in real time, and the timeliness and accuracy of monitoring are significantly improved.

Description

Technical Field

[0001] This application relates to the field of production data processing technology, and in particular to a method for monitoring the production quality of cold forging based on multi-source data fusion using an in-vehicle camera. Background Technology

[0002] As key sensors in advanced driver assistance systems (ADAS) and autonomous driving technologies, automotive cameras directly impact driving safety. Their internal precision metal components are complex in shape and require extremely high accuracy, typically manufactured using near-net-shape cold forging. However, cold forging dies inevitably wear under extreme pressure and high-frequency cyclic loads. This wear is not a simple linear accumulation but rather progresses through different stages: initial break-in, intermediate stabilization, and severe wear in the later stages. Minor geometric changes in the die cavity are directly transmitted to the forging, causing critical form and position tolerances to gradually deviate from the design center—a phenomenon known as tolerance drift. For optical components, this drift, or internal defects caused by flow anomalies, can lead to optical axis misalignment, seal failure, or assembly failure, severely impacting product quality.

[0003] Currently, the industry primarily relies on offline sampling inspections, such as those using coordinate measuring machines (CMMs) and statistical process control, to monitor product quality. The main drawback of these methods is their significant lag; they can only detect quality problems that have already occurred and cannot provide early warnings. Furthermore, sampling inspections cannot cover all products, potentially missing batch-to-batch fluctuations or sudden quality deterioration.

[0004] Existing data-driven prediction methods often treat die wear as a monotonically changing process, or assume that the influence coefficient of process parameters on tolerance drift is constant throughout the die's lifespan. However, this static assumption ignores the phased physical characteristics of die wear. In fact, from initial break-in to stable wear, and then to severe wear, the internal state of the die undergoes a qualitative change. At different wear stages, the sensitivity of the cold forging system to the same process parameter disturbances varies, leading to nonlinear changes in the rate and pattern of tolerance drift, and even the mechanism of internal defect occurrence. Existing models fail to capture this dynamic relationship dependent on the implicit wear state, causing their prediction accuracy to drop sharply when the wear state undergoes a critical transition, making it impossible to issue timely and accurate warnings. Summary of the Invention

[0005] To address the technical problems in existing technologies, such as lagging monitoring of internal defects in cold forging and difficulty in capturing the dynamic nonlinear characteristics of the process, which leads to inaccurate evaluation, this application provides a method for monitoring the production quality of cold forging based on multi-source data fusion using an onboard camera.

[0006] This application provides a method for monitoring the production quality of cold forging using an automotive camera based on multi-source data fusion. The method includes: collecting multi-source high-frequency sensor data from the cold forging process; preprocessing and extracting deep physical features from the multi-source high-frequency sensor data to obtain a real-time physical feature vector characterizing the dynamic process of the current stamping cycle; constructing a multi-sensor time-series ideal flow reference model characterizing the ideal material flow state for different working conditions based on defect-free sample data from historical production data; calculating the multi-dimensional flow deviation between the real-time physical feature vector and the ideal flow reference model for the corresponding working condition in real time; calculating a flow deviation risk index based on a defect pattern statistical model trained from historical defect data; and determining the internal defect risk level of the forging corresponding to the current stamping cycle based on the flow deviation risk index, thereby achieving real-time monitoring of the production quality of cold forging using an automotive camera.

[0007] This application constructs an ideal flow benchmark that adapts to working conditions and calculates the dynamic deviation between the current cycle and the benchmark in real time. Combined with a risk index based on the square of the Mahalanobis distance, it can sensitively capture material flow anomalies that indicate internal defects. This overcomes the limitations of the lag and static assumptions of traditional monitoring and achieves real-time and accurate assessment of the risk of internal defects in cold forging.

[0008] In one embodiment, the deep physical feature extraction includes the following features: filling stage features, including the main filling duration, the maximum instantaneous rise rate of the in-mold pressure, and the standard deviation of the peak time of the in-mold pressure at different locations; holding and unloading features, including the standard deviation of the holding stage pressure, the initial slope of the force-displacement unloading curve, and the elastic recovery work; and energy and work features, including plastic deformation work, total stamping work, and friction indicator.

[0009] By extracting deep features with clear physical meaning, rather than simply using raw curves or peaks, the dynamic details of the cold forging process can be characterized more robustly and interpretably, improving the sensitivity and accuracy of subsequent benchmark comparisons and deviation assessments.

[0010] In one embodiment, the construction of a multi-sensor time-series ideal flow reference mode characterizing the ideal material flow state includes: using the dynamic time warping centroid averaging method to iteratively align and average the normalized time-series curves corresponding to all ideal samples within the same working condition cluster; and using the centroid vector after iterative convergence as the ideal flow time-series reference mode of the sensor under the corresponding working condition.

[0011] In one embodiment, the different operating conditions are obtained by cluster analysis of the raw material batch number and billet initial temperature corresponding to the ideal sample data; the cluster analysis adopts the mean clustering algorithm.

[0012] In one embodiment, the multidimensional flow deviation includes calculating the difference between the real-time physical feature vector and the ideal flow reference pattern in at least one of the following: dynamic time regularization morphological distance; normalized area deviation; key point timing deviation; area exceeding the reference fluctuation band; and multi-sensor synchronization deviation.

[0013] By calculating multi-dimensional deviation characteristics, this invention can comprehensively capture the differences between actual flow and ideal benchmark from multiple perspectives, such as morphology, amplitude, key timing points, local details, and multi-sensor synergy. These differences are the direct basis for judging whether the flow is abnormal and whether there is a risk of defects.

[0014] In one embodiment, the flow deviation risk index satisfies the following relationship: ;in, This is the liquidity deviation risk index. The squared Mahalanobis distance from the multidimensional flow deviation vector to the center of the historical defect-free pattern is given. The squared nearest Mahalanobis distance from the multidimensional flow deviation vector to the center of all historical defect patterns. and The preset sensitivity tuning hyperparameters, To prevent division by zero of extremely small positive numbers, k is the index of the defect pattern category, and K is the total number of all defect pattern categories.

[0015] The construction of the flow deviation risk index enables risk scores to more accurately reflect the probability of actual defects occurring. Only when the deviation significantly deviates from the normal pattern and is closer in pattern to a certain known defect pattern in history will a high risk score be generated, thus enhancing the accuracy of risk assessment.

[0016] In one embodiment, the squared Mahalanobis distance satisfies the following relationship: ;in, For the current multidimensional flow deviation vector, For defect pattern category indexing, including defect-free pattern indexing and at least one defect pattern index , and The first one is obtained by training with historical data. The mean vector and regularized covariance matrix of the class pattern. This represents the transpose of a vector. This represents the inverse of the covariance matrix.

[0017] In one embodiment, the internal defect risk level is determined as follows: a low-risk threshold and a high-risk threshold are preset; if the flow deviation risk index is less than the low-risk threshold, it is determined to be of low risk level; if the flow deviation risk index is not less than the low-risk threshold and less than the high-risk threshold, it is determined to be of medium risk level, and the forging is marked for inclusion in enhanced sampling inspection; if the flow deviation risk index is not less than the high-risk threshold, it is determined to be of high risk, an alarm is triggered, and the forging is isolated.

[0018] By setting tiered thresholds and corresponding handling strategies, the abstract risk index is transformed into specific instructions for normal flow, enhanced sampling inspection, and immediate isolation operations, providing a clear decision-making basis for achieving real-time internal quality sorting and process intervention.

[0019] In one embodiment, the acquisition process of the multi-source high-frequency sensor data includes: embedding multiple high-frequency pressure sensors in the key flow path and final filling area within the mold cavity; installing high-precision dynamic force sensors and high-resolution displacement sensors on the main slide of the press to synchronously record force-displacement curves; monitoring the temperature near the surface of the mold cavity and before the billet enters the mold; and installing vibration or acoustic emission sensors in the fixed mold section.

[0020] The technical solution of this application has the following beneficial technical effects: This application can provide a risk level immediately after each stamping cycle by deeply analyzing readily available high-frequency sensor data patterns such as multi-point pressure and force-displacement, overcoming the lag and sampling limitations of traditional methods. At the same time, by constructing an ideal flow time series benchmark that is adaptive to working conditions and using in-depth indicators such as key point time series deviations, it greatly improves the sensitivity and identification ability of material flow anomalies, and can capture subtle dynamic pattern deviations that are easily overlooked by conventional peak monitoring and indicate internal defects.

[0021] Furthermore, by constructing a flow deviation risk index that integrates the degree to which the current flow deviation deviates from the normal state and the similarity of its pattern to historically known defect patterns, the risk score more accurately reflects the probability of actual defects occurring. Attached Figure Description

[0022] Figure 1 This is a flowchart of a method for monitoring the production quality of cold forging using an in-vehicle camera based on multi-source data fusion, according to an embodiment of this application. Detailed Implementation

[0023] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0024] Figure 1This is a flowchart illustrating a method for monitoring the production quality of cold forging using an onboard camera based on multi-source data fusion, according to an embodiment of this application. Figure 1 As shown, the method for monitoring the production quality of cold forging based on vehicle-mounted cameras using multi-source data fusion includes steps S101 to S104, which are described in detail below.

[0025] S101 collects multi-source high-frequency sensor data from the cold forging production process, and preprocesses and extracts deep physical features from the multi-source high-frequency sensor data to obtain a real-time physical feature vector characterizing the dynamic process of the current stamping cycle.

[0026] In one embodiment, in order to comprehensively and accurately capture the dynamic process information of the cold forging cycle and extract physical characteristics that are sensitive to changes in material flow state and mold condition, the sensor deployment and high-frequency data acquisition can be optimized first.

[0027] Specifically, multiple miniature high-frequency pressure sensors are embedded in key flow paths within the mold cavity, such as at the end of the flow channel, the confluence area, and the final filling region, or at the root of the thin wall and the bottom of the deep cavity. Examples of such sensors include piezoelectric thin-film sensors or fiber Bragg grating (FBG) sensors. The sampling frequency of these sensors needs to be set to the kHz level, for example, 10 kHz, to capture the details of transient pressure fluctuations during the material filling process.

[0028] A high-precision dynamic force sensor and a high-resolution displacement sensor are installed on the main slide of the press to synchronously record the complete force-displacement curve. The sampling frequency is consistent with that of the in-mold pressure sensor, for example, 10 kHz.

[0029] Thermocouples are embedded in key locations near the mold cavity surface, such as areas prone to adhesion, to monitor the mold's working temperature in real time. A high-speed infrared thermal imager or fiber optic thermometer is used to measure the surface temperature of the billet before it enters the mold.

[0030] A broadband accelerometer, such as 1 Hz - 20 kHz, and an acoustic emission sensor, such as 100 kHz - 1 MHz, are installed on the die holder or the fixed die section near the die cavity to simultaneously acquire vibration and acoustic emission signals during the stamping process, especially the high-pressure filling and unloading stages.

[0031] The manufacturing execution system or equipment PLC automatically records the precise cumulative number of times the die is used for each stamping, the furnace number / batch number of the raw materials, the type and parameters of the lubricant, and auxiliary data such as the temperature and humidity of the workshop environment.

[0032] Furthermore, the acquired multi-source high-frequency sensor data is synchronized, cleaned, and standardized. This includes using a hardware trigger signal, such as the threshold signal of a force sensor or displacement sensor generated when the punch contacts the billet, as the synchronization start point for all high-frequency data acquisition channels, ensuring that the relative deviation of the data in each channel on the time axis is controlled at the microsecond level.

[0033] Low-pass digital filters are applied to pressure and force signals to remove high-frequency electrical noise, band-pass filters are applied to vibration signals, and high-pass filters are applied to acoustic emission signals to remove low-frequency mechanical noise. Wavelet thresholding can be used to denoise all high-frequency signals to preserve abrupt changes.

[0034] All curves based on time t are uniformly resampled to a fixed number of points N through cubic spline interpolation or linear interpolation, for example, N=1024, to obtain vector representations of equal length. For force-displacement curves, data points with equal displacement intervals can be obtained first by interpolating based on displacement d.

[0035] Finally, each resampled curve vector is subjected to min-max normalization to scale its value range to the interval [0,1] or [-1,1], thereby eliminating the influence of different sensor ranges and signal amplitudes, making subsequent morphology-based comparisons more fair and effective.

[0036] In this optional embodiment, the original curve contains a large amount of redundant information and is sensitive to noise. Extracting deep physical key features that have clear physical meaning and can reflect the material flow state, filling degree, stress-strain history, friction conditions and potential damage events can more robustly and interpretably characterize the cold forging process, thereby obtaining a real-time physical feature vector characterizing the dynamic process of the current stamping cycle.

[0037] Specifically, for the characteristics of the filling stage, for example, these include the main filling start time, the main filling end time, the main filling duration, the maximum instantaneous rate of rise of all pressure curves during the main filling period, the precise time when each pressure curve reaches its peak pressure, the standard deviation of the peak pressure time for different pressures, the force and displacement values ​​at the yield point on the force-displacement curve, and the displacement value when the force-displacement curve reaches its plateau period.

[0038] For the characteristics of pressure holding and unloading, examples include the standard deviation of pressure of each pressure curve during the pressure holding stage, the slope of the initial part of the unloading segment of the force-displacement curve, the elastic recovery work represented by the area under the unloading curve F(d), and the hysteresis loss work represented by the area enclosed by the loading curve and the unloading curve.

[0039] Examples of energy and work characteristics include plastic deformation work, total stamping work, and friction indicators.

[0040] In an optional embodiment, advanced signal characteristics, for example, may include the root mean square value of the vibration signal, the kurtosis of the vibration signal, the energy of the acoustic emission signal, the ring count of the acoustic emission signal, the duration of the acoustic emission signal, etc. In this optional embodiment, all extracted features are finally combined into a high-dimensional physical feature vector to obtain a real-time physical feature vector characterizing the dynamic process of the current stamping cycle.

[0041] In this way, by acquiring multi-source high-frequency data and extracting deep physical features, the dynamic process information of the cold forging cycle can be captured comprehensively and with high fidelity, and transformed into robust and physically meaningful feature vectors, providing a high-quality data foundation for subsequent benchmark modeling and deviation assessment.

[0042] S102, based on defect-free sample data from historical production data, constructs a multi-sensor time-series ideal flow benchmark model to characterize the ideal material flow state under different working conditions.

[0043] In one embodiment, to accurately depict how the material should ideally flow when the mold is in a healthy state under specific operating conditions, it is necessary to first screen and label ideal samples. Specifically, from long-term production history data, samples that simultaneously meet the following conditions are rigorously screened as defect-free ideal samples: The corresponding forgings have undergone rigorous testing using methods such as ultrasonic scanning, industrial CT, or metallographic sectioning to confirm the absence of internal defects such as folds, porosity, and cracks. All key dimensional and positional tolerances of the corresponding forgings are within the specification requirements and close to the design center value, for example, within the range of mean ± 1σ (standard deviation). The cumulative number of uses of the corresponding molds is in the stable wear stage identified by the historical defect rate curve, for example, in the range of 20% to 80% of the total lifespan, avoiding the impact of initial break-in and severe wear at the end. The process parameters of the corresponding stamping cycle and its adjacent cycles, such as force, energy, and temperature, fluctuate little, with no obvious abnormal peaks or drifts.

[0044] In this optional embodiment, complete and preprocessed, normalized multi-sensor time-series curves and corresponding operating parameters corresponding to these ideal samples can be collected and clustered according to operating conditions. This is because the batch size of raw materials affects strength and plasticity, and the initial temperature of the billet affects flow stress and friction, which are the main external factors affecting the cold forging process. Different combinations will systematically change the shape of the ideal pressure curve. To improve the accuracy of the benchmark, the ideal samples need to be grouped according to operating conditions.

[0045] Specifically, key operating condition parameters are selected as clustering features. For example, the raw material batch number is thermally encoded, the initial surface temperature of the billet is taken as a continuous feature, and mean clustering algorithms such as K-Means clustering are applied. The optimal number of clusters is determined by using the profile coefficient or elbow method, thereby dividing the ideal sample into C operating condition clusters based on its operating condition feature vector.

[0046] Furthermore, the dynamic time warping centroid averaging method is employed to iteratively align and average the normalized time series curves corresponding to all ideal samples within the same operating condition cluster. Since the golden sample curves under the same operating condition still exhibit natural, minor time series fluctuations, direct arithmetic averaging would obscure the sharp characteristics of the curves. Dynamic time warping (DTW), however, can effectively align curves with nonlinear time series differences. The centroid averaging method, through iterative alignment and averaging, can calculate the morphological center of a set of sequences, making it an ideal choice for constructing representative benchmark patterns.

[0047] Specifically, for each operating condition cluster c, c∈{1,...,C}, the following procedure is performed for each sensor curve: Initialization: from cluster Randomly select a curve vector As the initial center of gravity .

[0048] Iterative process (r=0, 1, 2, ...): a. Alignment phase: for clusters Each curve vector in Calculate its relationship with the current center of gravity. Optimal DTW alignment path between The path It is a series of point pairs ,express The first in points and The first in Each point corresponds to one point.

[0049] b. Average phase: Calculate the new center of gravity The first on the new focus The value of each point Calculated as: all paths through their respective DTWs Mapping to centroid index The original curve points (in The arithmetic mean of ).

[0050] c. Convergence criterion: Compare the new centroid. With the old center of gravity The distance between them, for example, such as Euclidean distance, is used. If the distance is less than a preset convergence threshold or the maximum number of iterations is reached, the iteration stops.

[0051] Output: The centroid vector that finally converges. That is, the working condition The ideal flow timing reference mode of the sensor is shown below, and at the same time, in the averaging phase of the last iteration, for each point on the centroid... Calculate the standard deviation of all original curve point values ​​mapped to that point. Thus, the fluctuation range vector is obtained. The normal fluctuation band of the benchmark can be defined as follows: ,in It is a coefficient that controls the width of the fluctuation band. Its selection reflects the tolerance for normal fluctuations, and it is usually taken as 2 or 3.

[0052] Finally, the baseline mode vectors for all operating conditions c are... and fluctuation range vector Stored in a baseline pattern library for real-time retrieval.

[0053] In this way, a high-precision and condition-adaptive ideal flow reference model can be constructed, which accurately depicts the standard process shape under healthy conditions and provides a reliable reference scale for subsequent real-time deviation assessment.

[0054] S103 calculates the multi-dimensional flow deviation between the real-time physical feature vector and the ideal flow benchmark mode for the corresponding working condition in real time, and calculates the flow deviation risk index based on the defect mode statistical model trained on historical defect data.

[0055] In one embodiment, after each cold forging cycle, the difference between the current process and the ideal baseline is calculated in real time and converted into a risk assessment value for internal defects.

[0056] Specifically, the system acquires all raw sensor data for the current stamping cycle, performs preprocessing and normalization operations to obtain a normalized real-time curve vector; it also acquires the current operating condition parameters and loads the operating condition data by querying the reference mode library. Corresponding ideal reference mode and fluctuation range .

[0057] Furthermore, to comprehensively capture the differences between actual flow and ideal benchmark in terms of morphology, amplitude, key timing points, local details, and multi-sensor coordination, these differences serve as the basis for judging whether the flow is abnormal and whether there are potential defects or risks. For example, detailed flow deviation characteristics are calculated for each sensor and force-displacement curve: DTW morphological distance : ,in For the current stamping cycle; Normalized area bias ; Key timing deviations: ; Piecewise weighted RMSE: Dividing the curve into segments A key stage Calculate the root mean square error for each stage. The final shape deviation can be the weighted sum of the RMSE at each stage. The weight The sensitivity to defect formation at this stage can be set accordingly.

[0058] Area exceeding the baseline fluctuation band: Calculate the actual curve Beyond the benchmark fluctuation band The cumulative area of ​​a portion.

[0059] Multi-sensor synchronization deviation: calculation This refers to the maximum deviation between the relative peak times of each sensor and the baseline.

[0060] Similar deviation characteristics are calculated for the force-displacement curve. All calculated deviation characteristics form an initial deviation vector. A feature selection method can then be used to filter out the subset with the highest predictive power for defect risk, resulting in the final deviation vector. .

[0061] Furthermore, when calculating the flow deviation risk index, it is necessary to perform the calculation based on a defect pattern statistical model trained offline using historical defect data. This training includes data preparation and model parameter estimation. Specifically, the data preparation stage requires collecting a large amount of historical production data, including the final deviation vector for each sample and its internal defect labels determined through non-destructive testing, such as labels. Model parameter estimation is performed for each pattern category. Where k=0 represents the defect-free mode, and k>=1 represents different defect modes, calculate the mean vector and regularized covariance matrix of all sample bias vectors under this category. ,in It is the original calculated covariance matrix. Let be the regularized matrix. This is a small regularization parameter that adds a small positive value to the main diagonal of the original covariance matrix to ensure the matrix's invertibility and numerical stability. For example, 0.001 can be used to guarantee matrix invertibility. It is an identity matrix.

[0062] Furthermore, to comprehensively assess whether the current flow deviation significantly deviates from the normal pattern, yet is more closely similar in pattern to historical deviation patterns that led to internal defects, a flow deviation risk index can be calculated for evaluation. The flow deviation risk index satisfies the following relationship:

[0063] in, This is the liquidity deviation risk index. The squared Mahalanobis distance from the multidimensional flow deviation vector to the center of the historical defect-free pattern is given. The squared nearest Mahalanobis distance from the multidimensional flow deviation vector to the center of all historical defect patterns. and The preset sensitivity tuning hyperparameters, control function pairs Response speed to changes The larger, The earlier the function approaches 1, the faster the flow deviation risk index is affected by the degree of deviation from normal; hyperparameters To control the amplifying effect of the ratio term on the risk index, The larger the absolute value, the higher the risk index is magnified when the deviation is closer to the defect mode. To prevent division by zero of extremely small positive numbers, for example, the value is taken as... k is the index of the defect pattern category, and K is the total number of all defect pattern categories.

[0064] Specifically, the ratio term R = The flow deviation was evaluated as either a defect mode closer to the nearest R-value of 1 or a normal mode when R is not less than 1. Hyperparameters This is a negative value used to adjust sensitivity; for example, it takes the value -2 when R is less than 1. When R is not less than 1, The value is no greater than 1, which allows the exponential term to amplify the signal for situations close to defects.

[0065] The next item As a trigger gate The function will Mapping to the interval (0, 1) only when Large enough, meaning the deviation significantly deviates from the normal range, It's only close to 1; For example, the value is Used for control The sensitivity is such that if the deviation is close to a certain defect pattern but is itself very small. It will be very small, thus suppressing the final FDRI value.

[0066] Ultimately, FDRI combines relative proximity and absolute deviation. A high FDRI value indicates that the current flow deviation significantly deviates from the normal pattern, yet is also closer in pattern to a historically known defective pattern. It's worth noting that hyperparameters... , ϵ can be determined by optimizing on the validation set, for example by using methods such as grid search or gradient descent, to maximize the evaluation metric AUC (Area Under the Curve) or F1-score.

[0067] For example, suppose we monitor two types of defects: k=1 (folding) and k=2 (loosening), where k=0 represents no defects. Let the deviation vector be... It has been simplified to two dimensions. Furthermore, offline training yields defect-free mean values. , folded mean Loose mean To simplify the example, let all covariance matrices be... All are identity matrices hyperparameters , , .

[0068] The deviation vector is then calculated in real time by monitoring the current loop. Calculate the squares of each Mahalanobis distance: , , ;Sure , The closest folding pattern; finally, the ratio term of FDRI is calculated. exponent term , item , .

[0069] Thus, by constructing the FDRI index, the risk of current flow deviations can be comprehensively assessed in real time. Its calculation takes into account both the absolute degree of deviation from the normal and the relative similarity with known defect patterns, so the assessment results are more accurate and reliable.

[0070] S104 determines the risk level of internal defects in the forgings corresponding to the current stamping cycle based on the flow deviation risk index, so as to realize real-time monitoring of the cold forging production quality of vehicle cameras.

[0071] In one embodiment, receiver operating characteristic (ROC) curves can be plotted based on historical data of the calculated flow deviation risk index (FDRI) at different FDRI thresholds. The threshold point that best balances the true positive rate (correct detection of defects) and the false positive rate (false alarms) can be selected, and low-risk and high-risk thresholds can be set. For example, a high-risk threshold corresponding to a 95% true positive rate can be selected, while a low-risk threshold corresponding to a 10% false positive rate, thus setting the high-risk threshold to 100 and the low-risk threshold to 20.

[0072] In this optional embodiment, if the flow deviation risk index is less than the low-risk threshold of 20, it is determined to be at a low-risk level, indicating that the internal quality is qualified and the forging is being processed normally; if the flow deviation risk index is not less than the low-risk threshold of 20 and less than the high-risk threshold of 100, it is determined to be at a medium-risk level, and the forging is marked for inclusion in the enhanced sampling inspection. At the same time, the system records this event and related parameters for subsequent analysis; if the flow deviation risk index is not less than the high-risk threshold of 100, it is determined to be at a high risk, triggering an alarm and isolating the forging. At the same time, a warning message is pushed to the operator and engineer, including the FDRI value, the closest defect mode, and the main deviation characteristics that triggered the high risk.

[0073] Furthermore, to continuously optimize the accuracy of monitoring, the obtained defect pattern statistical model can be iteratively updated through knowledge accumulation and closed-loop feedback. Samples that are judged to be of medium or high risk and whose defect types have been verified by non-destructive testing are collected regularly to update the historical database. The updated database is then used for offline retraining, thereby recalculating the mean vector and covariance matrix of each pattern category, and re-optimizing the hyperparameters and risk thresholds in the FDRI relation, thus achieving continuous learning and performance improvement of the model.

[0074] In this way, through risk classification and closed-loop feedback, not only is real-time quality judgment and handling achieved, but also through continuous self-learning of the model and data mining, it can guide process optimization and predictive maintenance in reverse, thus forming a complete intelligent quality control closed loop.

[0075] It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this patent application shall be determined by the appended claims.

Claims

1. A method for monitoring the production quality of cold forging based on multi-source data fusion using vehicle-mounted cameras, characterized in that, include: Multi-source high-frequency sensor data is collected during the cold forging production process, and the multi-source high-frequency sensor data is preprocessed and deep physical feature extracted to obtain a real-time physical feature vector characterizing the dynamic process of the current stamping cycle. Based on defect-free sample data from historical production data, a multi-sensor time-series ideal flow benchmark model is constructed to characterize the flow state of ideal materials under different working conditions. The system calculates the multi-dimensional flow deviation between the real-time physical feature vector and the ideal flow reference mode for the corresponding working condition, and calculates the flow deviation risk index based on the defect pattern statistical model trained on historical defect data. The multi-dimensional flow deviation includes calculating the difference between the real-time physical feature vector and the ideal flow reference mode in at least one of the following aspects: Dynamic time-regulated morphological distance; Normalized area bias; Key timing deviations; Exceeding the area of ​​the benchmark fluctuation band; Multi-sensor synchronization deviation; The internal defect risk level of the forging corresponding to the current stamping cycle is determined based on the flow deviation risk index, so as to realize real-time monitoring of the cold forging production quality of vehicle-mounted cameras.

2. The method for monitoring the production quality of cold forging based on multi-source data fusion using an onboard camera, as described in claim 1, is characterized in that... The deep physical feature extraction includes the following features: Characteristics of the filling stage include the duration of the main filling, the maximum instantaneous rate of rise of the in-mold pressure, and the standard deviation of the peak time of the in-mold pressure at different locations; The characteristics of pressure holding and unloading include the standard deviation of pressure during the pressure holding stage, the initial slope of the force-displacement unloading curve, and the elastic recovery work. Energy and work characteristics, including plastic deformation work, total stamping work, and friction indicators.

3. The method for monitoring the production quality of cold forging based on multi-source data fusion using an onboard camera, as described in claim 1, is characterized in that... The construction of the multi-sensor time-series ideal flow reference model characterizing the ideal material flow state includes: The dynamic time warping centroid averaging method is adopted to iteratively align and average the normalized time series curves corresponding to all ideal samples within the same working condition cluster. The centroid vector after iterative convergence is used as the ideal flow timing reference mode of the sensor under the corresponding working conditions.

4. A method for monitoring the production quality of cold forging based on multi-source data fusion using an onboard camera, as described in claim 1 or 3, characterized in that... The different operating conditions are obtained by cluster analysis of the raw material batch number and billet initial temperature corresponding to the ideal sample data; the cluster analysis adopts the mean clustering algorithm.

5. The method for monitoring the production quality of cold forging based on multi-source data fusion using an onboard camera, as described in claim 1, is characterized in that... The flow deviation risk index satisfies the following relationship: in, This is the liquidity deviation risk index. The squared Mahalanobis distance from the multidimensional flow deviation vector to the center of the historical defect-free pattern is given. The squared nearest Mahalanobis distance from the multidimensional flow deviation vector to the center of all historical defect patterns. and The preset sensitivity tuning hyperparameters, To prevent division by zero of extremely small positive numbers, k is the index of the defect pattern category, and K is the total number of all defect pattern categories.

6. The method for monitoring the production quality of cold forging based on multi-source data fusion using an onboard camera, as described in claim 5, is characterized in that... The squared Mahalanobis distance satisfies the following relationship: in, For the current multidimensional flow deviation vector, For defect pattern category indexing, including defect-free pattern indexing and at least one defect pattern index , and The first one is obtained by training with historical data. The mean vector and regularized covariance matrix of the class pattern. This represents the transpose of a vector. This represents the inverse of the covariance matrix.

7. The method for monitoring the production quality of cold forging based on multi-source data fusion using an onboard camera, as described in claim 1, is characterized in that... The method for determining the risk level of the internal defect is as follows: Pre-set low-risk and high-risk thresholds; If the flow deviation risk index is less than the low risk threshold, it is determined to be a low risk level; If the flow deviation risk index is not less than the low risk threshold and less than the high risk threshold, it is judged as medium risk level and the forging is marked for inclusion in the enhanced sampling inspection. If the flow deviation risk index is not less than the high risk threshold, it is judged as high risk, triggering an alarm and isolating the forging.

8. The method for monitoring the production quality of cold forging based on multi-source data fusion using an onboard camera, as described in claim 1, is characterized in that... The data acquisition process of the multi-source high-frequency sensor includes: Multiple high-frequency pressure sensors are embedded in the key flow paths and final filling areas within the mold cavity; A high-precision dynamic force sensor and a high-resolution displacement sensor are installed on the main slide of the press to synchronously record the force-displacement curve. Monitor the temperature near the mold cavity surface and before the blank is placed into the mold; Vibration or acoustic emission sensors are installed in the fixed mold section.

Images