Intelligent inhibition and parameter optimization method and system for casting process defects

Abstract

The present application relates to the technical field of data processing, and more particularly to a method and system for intelligent inhibition and parameter optimization of casting process defects. The method comprises: dividing a standardized time series data sequence by sliding a sliding time window, extracting the deviation of each sensing channel signal from the normal working condition, and gathering it into a multi-dimensional feature vector of each sliding time window; inputting the multi-dimensional feature vector of each sliding time window into a judgment model, constructing a feature history buffer zone by a plurality of continuous normal windows before the current sliding time window, estimating a dynamic judgment threshold based on the feature distribution in the buffer zone, and outputting an abnormal flag; after confirming the effectiveness of the continuous multiple abnormal flags, generating a hierarchical process adjustment instruction according to the deviation of the multi-dimensional feature vector from the normal baseline and issuing it for execution. The present application can effectively distinguish between normal physical fluctuations and real defect signs in the casting process, simultaneously inhibit false positives and false negatives under the condition of working condition drift, and significantly improve the yield of casting products and process consistency.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for intelligent suppression and parameter optimization of casting process defects. Background Technology

[0002] Casting is a key manufacturing technology that involves injecting liquid materials (such as molten metal, resin, ceramic slurry, etc.) into a predetermined mold cavity and allowing them to solidify and take shape. It is widely used in aerospace precision casting, electronic component packaging, and the manufacturing of insulation for large power equipment. With increasingly stringent requirements for the internal quality and mechanical properties of industrial products, internal defects such as air bubble entrapment, shrinkage cavities, cold shuts, and incomplete filling during the casting process have become core bottlenecks restricting product reliability and yield. The formation of casting defects is strongly coupled with multiple process parameters such as filling speed, casting temperature, holding pressure, and mold preheating temperature. Moreover, the incubation and development of defects exhibit significant time-varying nonlinearity and strong randomness. Traditional process debugging methods that rely on manual experience are no longer sufficient to meet the demands of modern manufacturing for consistency and high yield.

[0003] Existing technologies have proposed solutions for monitoring the quality of the casting process through real-time sensor data. Typical solutions involve placing thermocouples and pressure sensors at key locations in the mold (such as near the gate, at the end of the cavity, and in thicker sections) to collect temperature change curves and dynamic pressure data throughout the casting process. Fixed upper and lower alarm thresholds are set for key physical quantities such as the rate of temperature drop, peak pressure, and pressure holding time during the holding phase. During the casting process, when the real-time monitoring value of any sensor channel exceeds the corresponding preset threshold, the system triggers an alarm or executes preset protective actions such as emergency pressure relief or stopping the casting process, in order to achieve the purpose of defect prevention and control.

[0004] The aforementioned monitoring scheme based on fixed thresholds has the following technical problems under actual dynamic casting conditions: the fixed alarm threshold cannot effectively distinguish between normal physical fluctuations and signs of actual defect incubation during the casting process. This is because the ambient temperature, batch material composition (such as fluctuations in the content of trace elements in the melt leading to deviations in solidification characteristics), and mold surface condition (such as changes in the heat exchange coefficient caused by coating wear due to continuous production) in actual casting production are constantly in a state of slow drift. These factors directly cause the normal baseline of key process parameters such as temperature decay characteristics and pressure build-up time to fluctuate by a non-negligible amplitude. Therefore, when the baseline drift exceeds the threshold margin, the fixed threshold generates a large number of false alarms, interfering with normal production. When the amplitude of actual defect signs (such as the pressure drop corresponding to shrinkage precursors) is less than the deviation between the fixed threshold and the current baseline due to the shift in operating conditions, the actual defect signs are missed, making it impossible to achieve early and reliable identification and effective suppression of defects. Summary of the Invention

[0005] To address the technical problem that fixed alarm thresholds cannot adapt to baseline drift in casting conditions, making it difficult to distinguish between normal process fluctuations and early defect characteristics, thus frequently resulting in false alarms and missed defect reports, and failing to achieve early identification and control of casting defects, this invention provides a method and system for intelligent suppression and parameter optimization of casting process defects.

[0006] In a first aspect, the present invention provides a method for intelligent suppression and parameter optimization of casting process defects, employing the following technical solution:

[0007] A method for intelligent suppression and parameter optimization of casting process defects, including the following steps: The original timing data sequences of multiple sensor channels arranged at various measuring points of the mold during the casting process are obtained, and the original timing data sequences are standardized to obtain the standardized timing data sequences of each sensor channel. The data is divided into a standardized time series by sliding a preset sliding time window. Statistical feature calculations are performed on the data of each sensor channel within each sliding time window. Based on the calculation results, the degree of deviation of each sensor channel signal from the normal operating condition is extracted, and the degree of deviation of each sensor channel is aggregated into a multi-dimensional feature vector of each sliding time window. The multidimensional feature vectors of each sliding time window are sequentially input into the judgment model. The judgment model uses the multidimensional feature vectors of multiple sliding time windows that were judged to be in a normal state before the current sliding time window to form a feature history buffer. The dynamic discrimination threshold is estimated based on the distribution of feature vectors in the feature history buffer. The deviation of the multidimensional feature vector of the current sliding time window from the dynamic discrimination threshold is used as the basis for the anomaly judgment of the current sliding time window, and an anomaly flag is output. After confirming the validity of the abnormal state based on the determination results of multiple consecutive abnormal indicators, a corresponding graded process adjustment instruction is generated according to the degree of deviation of the multidimensional feature vector from the normal baseline.

[0008] The ingenuity of this invention lies in the standardization and sliding window analysis of the casting timing data from multiple measurement points and multiple sensing channels of the mold. Multidimensional features are continuously extracted from the deviations of the signal from normal operating conditions, and the feature distribution of recent consecutive normal windows is used to dynamically determine the discrimination threshold. This allows the anomaly judgment benchmark to adjust to the slow operating condition drift caused by non-defect factors such as ambient temperature, material batch, and mold condition. This more effectively distinguishes between normal physical fluctuations and early signs of defects such as porosity, shrinkage cavities, cold shuts, or incomplete filling. Simultaneously, continuous anomaly confirmation suppresses erroneous actions caused by instantaneous noise, and graded process adjustment instructions are generated based on the degree of anomaly deviation, improving the early suppression capability of casting defects and enhancing product yield and batch process consistency.

[0009] Preferably, the standardized time-series data sequences of each sensing channel are obtained in the following manner: The original sequences of each sensor channel are obtained by collecting data at a preset sampling frequency until the end of the casting process. Any sensor channel is selected as a reference sensor channel, and the time delay of each sensor channel relative to the reference sensor channel is obtained through cross-correlation time delay estimation. Time-shift compensation is then applied to the original time-series data sequences of each sensor channel based on the time delay to eliminate time deviations between the sensor channels. The time-shift compensated data of each sensor channel is then smoothed and denoised using a sliding filter window to obtain the filtered time-series data sequences of each sensor channel. The global mean and standard deviation of the filtered time-series data sequences of each sensor channel are calculated. The data value at each sampling time is subtracted from the global mean and then divided by the standard deviation to obtain the standardized time-series data sequences of each sensor channel.

[0010] Preferably, the multidimensional feature vector is obtained in the following manner: The standardized time-series data is divided by a preset sliding time window. For each sliding time window, the mean and standard deviation of the data from each sensing channel within the window, as well as the linear trend coefficient of the fitted line, are calculated. The instantaneous deviation is obtained by the deviation of the last time value within the window from the mean within the window; the trend deviation is obtained by the difference between the linear trend coefficient and the historical normal reference trend coefficient; and the fluctuation distortion is obtained by the relative deviation of the standard deviation from the historical normal reference standard deviation. The three deviations of each channel within the same window are arranged into three vectors, and the norm of each vector is calculated as an anomaly intensity scalar for each dimension and concatenated into a multidimensional feature vector.

[0011] By simultaneously acquiring the instantaneous amplitude deviation, signal evolution trend deviation, and fluctuation amplitude distortion of the data at the end of the window, and by converging the responses of multiple sensing channels on the same deviation dimension into the overall abnormal intensity, the evaluation of the casting state is no longer limited to whether it exceeds the absolute threshold at a certain moment.

[0012] Preferably, the method for obtaining the historical normal reference trend coefficient and the historical normal reference standard deviation is as follows: A batch parameter A is preset, and at least A batches of qualified casting batches that are continuously produced and confirmed to be free of internal defects by non-destructive testing are selected. For each qualified batch, the standardized time-series data sequence of each sensor channel throughout the entire casting process is obtained, and the linear trend coefficient and data standard deviation of each sensor channel throughout the entire casting process are calculated. The mean of the linear trend coefficients of the same sensor channel in all qualified batches is used as the historical normal reference trend coefficient of that sensor channel, and the mean of the data standard deviations of the same sensor channel in all qualified batches is used as the historical normal reference standard deviation of that sensor channel.

[0013] By establishing reference values ​​for the normal trend and normal fluctuation of each sensing channel based on consecutive qualified casting batches confirmed by non-destructive testing to be free of internal defects, the benchmark for comparing abnormal characteristics is supported by actual product quality results.

[0014] Preferably, the determination model is constructed in the following ways: A feature history buffer is constructed using the multidimensional feature vectors of multiple consecutive sliding time windows that were judged to be in a normal state before the current sliding time window. For each dimension of the multidimensional feature vector of the current sliding time window, the deviation of the value of that dimension relative to the mean and standard deviation of the corresponding dimension in the feature history buffer is calculated, and the deviations of each dimension are concatenated into a deviation vector. The square root of the sum of the squares of each dimension of the deviation vector is used to obtain the comprehensive anomaly index of the current sliding time window. The mean and standard deviation of the comprehensive anomaly index of all historical sliding time windows in the feature history buffer are calculated, and the sum of the mean and the standard deviation by a preset multiple is used as the dynamic discrimination threshold of the current sliding time window.

[0015] The dynamic discrimination threshold is adaptively adjusted based on the normal center position and fluctuation range of recent casting conditions. This avoids frequent overshooting of normal data after the baseline of the working condition moves slowly, and maintains sensitivity to minor defect signs when the normal fluctuation is small.

[0016] Preferably, the feature history buffer is updated as follows: if the current sliding time window is determined to be normal, its multidimensional feature vector is stored in the buffer and the earliest stored feature vector is removed; if the current sliding time window is determined to be abnormal, the buffer is not updated.

[0017] Preferably, when the number of historical sliding time windows stored in the feature history buffer does not reach the preset capacity, a preset initial threshold is used as the discrimination threshold for the current sliding time window; the initial threshold is obtained by taking the product of the maximum value and the safety factor of the comprehensive anomaly index corresponding to the initial sliding time window of no less than batch A of qualified casting batches within the process debugging cycle.

[0018] Preferably, the step of confirming the validity of the abnormal state based on the determination results of multiple consecutive abnormal flags includes: The number of consecutive confirmation windows is preset. If the abnormal flags of the previous consecutive confirmation windows are all state values ​​that indicate an abnormality, the abnormal state of the current sliding window is confirmed to be valid. If at least one abnormal flag representing a normal state value exists in several consecutive confirmation windows preceding the current sliding time window, then the current abnormality is determined to be an occasional disturbance, and no action is triggered.

[0019] Preferably, generating the corresponding graded process adjustment instruction includes: The sum of the relative deviations of the values ​​of each dimension in the multidimensional feature vector of the current sliding time window from the mean of each dimension in the feature history buffer is used as the comprehensive degree of abnormal deviation. A first-level threshold and a second-level threshold are preset. When the deviation does not exceed the first-level threshold, a micro-adjustment instruction containing a first adjustment amount is generated. When the deviation exceeds the first-level threshold but does not exceed the second-level threshold, a medium-level adjustment instruction containing a second adjustment amount is generated. When the deviation exceeds the second-level threshold, a large-scale adjustment instruction containing a third adjustment amount is generated. The second adjustment amount is greater than the first adjustment amount, the third adjustment amount is greater than the second adjustment amount, and the large-scale adjustment instruction also includes triggering a pouring termination instruction if the continuous judgment after execution is still abnormal.

[0020] It can reduce excessive intervention when there are mild anomalies, provide sufficient corrective strength when there are moderate to severe anomalies, and implement safety protection in a timely manner when adjustments are ineffective, thereby taking into account defect suppression effect, process stability, material loss control and equipment operation safety.

[0021] Secondly, this invention provides an intelligent system for suppressing defects and optimizing parameters in the casting process, employing the following technical solution: A smart system for suppressing and optimizing casting process defects includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the aforementioned smart system for suppressing and optimizing casting process defects.

[0022] By adopting the above technical solution, the above-mentioned intelligent suppression and parameter optimization method for casting process defects is generated into a computer program and stored in a memory so that it can be loaded and executed by a processor. In this way, a terminal device can be made based on the memory and the processor for convenient use.

[0023] This invention has the following technical advantages: It simultaneously characterizes the deviation of each channel signal from the normal baseline from three dimensions—amplitude shift, trend evolution, and fluctuation pattern—using a sliding time window as the unit. This overcomes the limitation that a single absolute threshold cannot capture early dynamic anomalies of defects, allowing defects such as shrinkage and porosity to be detected during the incubation stage. It utilizes a feature history buffer to track operating condition drift in real time, enabling the judgment threshold to adaptively update with the slow changes in non-defect factors such as ambient temperature and batch material composition. This fundamentally solves the problem of baseline mismatch leading to numerous false alarms or masking defect signs within changes in the qualified baseline when the fixed threshold is used during operating condition drift, resulting in missed alarms. Furthermore, it filters out occasional disturbances through a continuous confirmation mechanism to avoid erroneous actions, and then implements graded process intervention based on the degree of deviation, achieving a complete closed loop from anomaly detection to graded adaptive suppression. Through the synergistic effect of the above-mentioned technical features, the present invention can effectively distinguish between normal physical fluctuations and real defect signs in complex dynamic casting scenarios where working conditions are constantly drifting, while suppressing false alarms and missed alarms, and automatically implementing graded process interventions during the defect incubation stage, ultimately significantly improving the online detection rate of casting defects, reducing the false alarm rate, and improving product yield and process consistency. Attached Figure Description

[0024] Figure 1 This is a flowchart of the intelligent suppression and parameter optimization method for casting process defects in this embodiment of the invention; Figure 2 This is a comparison chart showing the qualified status of products in consecutive production batches. Detailed Implementation

[0025] 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 some embodiments of the present invention, but not all embodiments.

[0026] This invention discloses a method for intelligent suppression and parameter optimization of casting process defects, referring to... Figure 1 This includes steps S1-S4: S1: Obtain the original timing data sequence of multiple sensor channels arranged at various measuring points of the mold during the casting process, and perform timing alignment, noise removal and normalization transformation on each of the original timing data sequences to obtain the standardized timing data sequence of each sensor channel.

[0027] It should be noted that the raw signals collected by the sensing units (such as thermocouples, pressure sensors, and velocity sensors) at various measuring points in the mold during the casting process have the following problems: different physical quantities (temperature, pressure, velocity) have different dimensions and amplitude levels, making direct comparison or joint analysis impossible; furthermore, electromagnetic interference sources such as the start-up and shutdown of high-power equipment and harmonics from frequency converters will superimpose high-frequency noise onto the sensing signals, masking the actual physical process changes; therefore, if the raw signals are directly used for subsequent defect identification, the difference in dimensions will make the different signals incomparable on a numerical scale, and noise interference may be misjudged as abnormal signs; therefore, step S1 transforms the raw signals of multiple sensing channels into dimensionless standardized time-series data sequences by performing time-series alignment, noise removal, and normalization transformation, making the signals of each sensing channel comparable on an amplitude scale, while filtering out high-frequency measurement noise.

[0028] In this embodiment of the invention, synchronous data acquisition is initiated when the pouring start signal is triggered, within the mold preset... Sensing units are arranged at key measurement points, including but not limited to K-type thermocouples for acquiring temperature signals, piezoresistive pressure sensors for acquiring pressure signals, and speed sensors for acquiring filling speed signals. All sensing channels operate at a preset raw sampling frequency. Synchronous data acquisition continues until the pouring end signal is triggered, obtaining the raw time-series data sequence acquired by each sensor channel; the raw sampling frequency is preset. In other embodiments, implementers may preset the original sampling frequency according to the specific implementation method. The value; It should be noted that, considering the differences in physical response time constants and signal transmission path delays among different sensing channels, there is a time shift between the sensing channels. Therefore, it is necessary to use the cross-correlation time delay estimation method to correct the time shift of the original data sequence of each sensing channel.

[0029] In this embodiment of the invention, an arbitrary sensing channel is selected as the reference sensing channel, and the cross-correlation function between the i-th sensing channel and the reference sensing channel is calculated. ,in, This represents the original time-series data sequence of the i-th sensing channel; The original timing data sequence of the reference sensor channel; This indicates that the original timing data sequence of the reference sensor channel will be shifted backward. 1 sampling point; the cross-correlation function The corresponding value when the maximum value is obtained The value is used as the relative time delay of the i-th sensing channel. ,like This indicates that the timing of the i-th sensing channel is ahead of the reference sensing channel, and the original timing data of the i-th sensing channel is shifted backward. Compensation was completed at each sampling point; if This indicates that the timing of the signal from the i-th sensor channel lags behind that of the reference sensor channel, and the original timing data of the i-th sensor channel is shifted forward. Each sampling point completes compensation, eliminating the relative time deviation between each sensing channel; Furthermore, to avoid measurement noise in the timing data after all sensor channels are aligned, the aligned timing data sequences of each sensor channel are denoised; specifically, taking the first... Taking the aligned timing data sequence of one sensing channel as an example, the filter window length is set to... For the first Sampling times in the aligned timing data sequence of each sensing channel Based on sampling time Centered on the window, extract symmetrical continuous segments on both sides of that moment. 1 data point, as the sampling time The filtering window will include the sampling time. The average value of all data points within the filtering window is used as the sampling time. The smoothed filter output value, for the first Smoothing is performed at each sampling time in the aligned timing data sequence of the nth sensing channel to obtain the nth... The filtered timing data sequence of each sensing channel; in this embodiment of the invention, a preset filtering window length is used. for In other embodiments, the implementer may pre-set a sampling point according to the specific implementation method. The value; Furthermore, to eliminate inconsistencies in dimensions and amplitude scales caused by differences in sensor unit type, installation location, and measurement range among various sensor channels, the filtered data of each sensor channel undergoes global normalization processing; specifically, for the first... The filtered timing data sequence of the nth sensing channel is used to calculate the nth... The mean of all data points for each sensor channel and standard deviation , obtain the The first sensing channel Normalized data values ​​at each sampling time point: ; in, For the first The first sensing channel Normalized data values ​​at each sampling time point; For the first The mean of all data points in each sensor channel; For the first The standard deviation of all data points in each sensor channel; it should be noted that the absolute amplitude of each sensor channel is uniformly transformed into a dimensionless relative quantity centered at zero with a unit standard deviation, so that different physical quantities (temperature, pressure or velocity) are comparable on a numerical scale. At the same time, since the mean and standard deviation are calculated based on the entire casting process, the relative change pattern and temporal evolution information of each sensor channel on the entire casting time axis are completely preserved.

[0030] The first The normalized data values ​​of all sampling times of the _th sensing channel constitute the _th The standardized time-series data sequences of each sensor channel are used to form a standardized multi-sensor channel time-series data matrix. Each row of the matrix corresponds to the standardized time-series data sequence of one sensor channel, and each column corresponds to the normalized data values ​​of all sensor channels at the same sampling time.

[0031] S2: Divide the standardized time-series data sequence by sliding a preset sliding time window, perform statistical feature calculations on the data of each sensor channel within each sliding time window, extract the degree of deviation of each sensor channel signal relative to the normal operating condition based on the calculation results, and aggregate the degree of deviation of each sensor channel into a multi-dimensional feature vector of each sliding time window.

[0032] It should be noted that although step S1 eliminates the dimensional differences between the various sensing channels through global normalization, making the signals of different physical quantities such as temperature and pressure comparable in amplitude scale, it still cannot be directly used for defect identification. This is because the normal signal baseline during the casting process will slowly drift with non-defect factors such as ambient temperature, batch material composition, and mold wear. This drift is reflected in the normalized time series data as an overall rise or fall of the signals of each sensing channel, a systematic shift in the rate of change, and a change in the amplitude of fluctuation. If the absolute value or rate of change after normalization is directly compared with a fixed threshold, it is impossible to distinguish between normal changes caused by working condition drift and abnormal changes caused by defect incubation. Therefore, this step uses a sliding time window as a unit to compare the signal characteristics of each sensing channel within the sliding time window with the channel's own mean value within that window and the statistical reference value of the channel under historical normal operating conditions. Multi-dimensional quantitative indicators that can reflect the degree of deviation from normal operating conditions are extracted from three dimensions: amplitude, trend, and fluctuation pattern.

[0033] In this embodiment of the invention, the length of the preset sliding time window is... There are 10 sampling points, and the preset window sliding step size is 10 ... One sampling point; Among them, the sliding time window length and step length The parameters can be preset based on the total duration of the casting process and the desired anomaly detection response speed. The window length must be sufficient to cover the minimum feature cycle of the main physical changes during the casting process, and the step size must ensure sufficient overlap between adjacent windows to maintain the smooth and continuous evolution of features. In this embodiment of the invention, the preset sliding time window length... For 500 sampling points, the window step size is... 100 sampling points; The sliding time window is divided on the standardized time-series data of each sensor channel to obtain several sliding time windows for each sensor channel; the mean of the data in each sliding time window of each sensor channel is used as the amplitude center position in each sliding time window of each sensor channel, and the standard deviation of all data in each sliding time window of each sensor channel is obtained. By performing univariate linear regression on the data in each sliding time window of each sensing channel, the fitted straight line in each sliding time window of each sensing channel is obtained. The slope of the fitted straight line is used as the linear trend coefficient of each sliding time window of each sensing channel. The sign and magnitude of the linear trend coefficient reflect whether the sensing channel signal is rising or falling within the time window and its rate of change. The physical meaning of the linear trend coefficient is that the physical evolution trends such as the pressure rise rate during the casting and filling stage, the pressure stability during the holding stage, and the temperature drop gradient during the cooling stage will all deviate from the normal operating conditions due to the generation of defects such as internal shrinkage or porosity. Therefore, incorporating trend information into feature extraction helps to capture early dynamic anomalies of defects. It should be noted that the mean, standard deviation and linear trend coefficient mentioned above are all conventional statistics in the field, and their specific calculation methods are well known to those skilled in the art. In the embodiments of the present invention, they will not be described in detail. In this embodiment of the invention, the first... The first sensing channel Instantaneous deviation of each sliding time window: ; in, Indicates the first The first sensing channel The instantaneous deviation of each sliding time window; Indicates the first The first sensing channel The normalized data value at the last sampling moment in a sliding time window; Indicates the first The first sensing channel The mean of data within a sliding time window; Indicates the first The first sensing channel The standard deviation of all data within a sliding time window; This represents a very small positive number, used to prevent the denominator from being zero. In this embodiment of the invention, a preset value is used. ; A normal fluctuation range for the signal is constructed using data from adjacent time periods covered by a sliding time window. The normalized data value at the last sampling moment is placed within the sliding time window for positioning. The larger the value, the more significant the deviation of the data at the last sampling time from the recent normal fluctuation range, and the positive or negative sign indicates the direction of the deviation; Furthermore, the linear trend coefficient of each sliding time window of each sensor channel is compared with the reference trend coefficient established for each sensor channel under historical normal operating conditions to obtain the trend deviation of each sliding time window of each sensor channel: ; in, Indicates the first The first sensing channel Trend deviation of each sliding time window; Indicates the first The first sensing channel Linear trend coefficients for each sliding time window; For the first The reference trend coefficient established by each sensing channel under historical normal operating conditions; even if the instantaneous amplitude at the current sampling moment has not exceeded the normal fluctuation range, the rate of change of the signal has deviated from the typical evolution trajectory under qualified operating conditions. This trend precursor often appears earlier than the amplitude exceeding the limit. Therefore, the trend deviation index has higher sensitivity to early defects. Furthermore, the second-order moment distortion of each sliding time window for each sensing channel is calculated to measure the change in the dispersion of the data distribution within the window relative to the normal expectation: ; in, Indicates the first The first sensing channel The second-order moment distortion of a sliding time window; Indicates the first The first sensing channel The standard deviation of all data within a sliding time window; Indicates the first Reference values ​​of the sample standard deviation of each sensor channel under historical normal operating conditions; When unstable flow or localized abnormal solidification occurs during the casting process, the fluctuation amplitude of the sensor signal often increases or decreases significantly compared to normal operating conditions. This is reflected in the relative deviation of the standard deviation within the window. The greater the deviation of the second-order moment distortion from zero, the more severe the distortion of the signal fluctuation amplitude within the window relative to the historical normal level. When the second-order moment distortion is greater than 0, it indicates that unstable flow or localized abnormal solidification during the casting process causes large-scale random fluctuations in the signal. When the second-order moment distortion is less than 0, it indicates that changes in the state of the medium around the measuring point cause the signal to tend to flatten and the fluctuation amplitude to decrease abnormally. Regardless of whether the deviation is in the positive or negative direction, it represents a substantial change in the fluctuation pattern under normal operating conditions and can be used as a basis for anomaly identification. It should be noted that a batch parameter A=30 is preset, as mentioned above. and The method for obtaining them is as follows: Select continuously produced materials that have been confirmed to have no internal defects by subsequent non-destructive testing (such as X-ray inspection or ultrasonic testing). One qualified casting batch The value of must satisfy the statistical condition of a large sample, requiring . The value of is not less than batch A. In this embodiment of the invention, a qualified casting batch is preset. In other embodiments, implementers may pre-determine specific implementation methods. The value; For each qualified batch, obtain the standardized time-series data sequence of all sensor channels throughout the entire casting process. Based on the standardized time-series data sequence of each sensor channel throughout the entire casting process, obtain the linear trend coefficient and data standard deviation of each sensor channel throughout the entire casting process. All qualified batches The mean of the linear trend coefficients of each sensing channel throughout the entire casting process is used as the first... The reference trend coefficient established by each sensor channel under historical normal operating conditions; the first of all qualified batches The mean of the standard deviations of the data from each sensor channel throughout the entire casting process is used as the first... The reference standard deviation of each sensor channel under historical normal operating conditions; it should be noted that in subsequent mass production, every 10 new batches that have passed inspection and confirmation will be added to the qualified casting batches, and the above values ​​will be recalculated. and The value of is determined to achieve continuous calibration of the reference baseline as the process evolves over a long period.

[0034] It should be noted that the above indicators describe the deviation of the current operating condition from the normal baseline from three different perspectives: amplitude, trend, and volatility. For each sliding time window, the deviation indicators of the above sensor channels are aggregated and fused to construct a multi-dimensional feature vector.

[0035] In this embodiment of the invention, the first of all sensing channels is... The instantaneous deviation of each sliding time window is arranged in order of the sensing channels to obtain the instantaneous deviation vector; the instantaneous deviation vector is obtained by arranging the first and second sliding time windows of all sensing channels. The trend deviation of each sliding time window is arranged according to the sensor channel order to obtain the trend deviation vector; the trend deviation vector is obtained by arranging the trend deviation of each sensor channel in sequence. The second-order moment distortion degrees of each sliding time window are arranged in order of the sensing channels to obtain the fluctuation distortion vector; the distortion of each vector is calculated separately. The norm, as a scalar of the overall anomaly strength of each vector, is the norm of the vector. The method for calculating the norm is a well-known technique in the art, and will not be elaborated further in this embodiment of the invention; the instantaneous deviation vector, trend deviation vector, and fluctuation distortion vector are concatenated to form the first norm. Multidimensional feature vectors of a sliding time window; The aforementioned overall anomaly intensity scalar comprehensively characterizes the overall deviation of the sliding time window from the normal operating condition boundary from three dimensions: instantaneous amplitude shift, trend evolution shift, and fluctuation pattern shift of multiple sensor channels. The deviation of a single index of a single sensor channel may originate from local random disturbances, but when multiple sensor channels show consistent deviations in multiple statistical dimensions at the same time, the cooperative correlation structure of the physical field inside the casting has undergone substantial changes. This cooperative change is highly related to physical mechanisms such as the obstruction of heat transfer paths in the initial stage of volumetric defects such as shrinkage porosity and pores.

[0036] S3: Construct a feature history buffer using the multidimensional feature vectors of multiple consecutive sliding time windows that were judged to be in a normal state before the current sliding time window. Estimate the dynamic discrimination threshold under the current working condition based on the distribution of feature vectors in the feature history buffer. Use the deviation of the multidimensional feature vector of the current sliding time window from the dynamic discrimination threshold as the basis for judging the anomaly of the current sliding time window and output the anomaly flag.

[0037] It should be noted that during the casting process, the baseline deviation of normal operating conditions caused by factors such as ambient temperature drift, batch material differences, and slow changes in mold condition often has an amplitude on the same order of magnitude as the signal deviation in the early stages of defect incubation. If a fixed threshold is used for anomaly detection, a threshold that is too narrow will lead to a large number of false alarms, while a threshold that is too wide will lead to missed defects. Therefore, this step requires the construction of a mechanism that can adaptively adjust the dynamic threshold in real time according to changes in the casting conditions, ensuring that normal fluctuations under the current conditions always fall within an acceptable boundary range, while deviations exceeding this dynamic boundary are identified as abnormal signs with defect risk. Based on this, this step uses the multidimensional feature vector sequence of each sliding time window as input, and sequentially performs dynamic baseline estimation based on historical feature buffers, preliminary judgment of abnormal signs based on relative deviation, and adaptive updating of the dynamic threshold, thereby providing a dynamically adjusted judgment basis for the subsequent final decision.

[0038] In this embodiment of the invention, a fixed length is constructed. A feature history buffer, which stores consecutive data prior to the current sliding time window. A multidimensional feature vector of a historical normal sliding time window is used to characterize the distribution of normal operating status under recent pouring conditions; in this embodiment of the invention, a fixed length of the preset historical feature buffer is used. In other embodiments, the implementer may preset a fixed length value according to the specific implementation method; The buffer is maintained as follows: after any sliding time window completes the anomaly detection and outputs an anomaly flag of 0, the sliding time window is a normal sliding time window. In this case, the multidimensional feature vector of the sliding time window is stored in the feature history buffer, and the earliest stored multidimensional feature vector is removed to keep the capacity unchanged. If the sliding time window outputs an anomaly flag of 1, the sliding time window is an abnormal sliding time window, and the feature history buffer is not updated. For each dimension of the multidimensional feature vector in the current sliding time window, it is compared with the feature history buffer. By comparing the corresponding dimensions of the multidimensional feature vectors, the deviation of each dimension in the multidimensional feature vector of the current sliding time window is obtained: ; In the formula, The th feature vector in the current sliding time window Deviation in each dimension; The th feature vector in the current sliding time window The values ​​that can be taken in each dimension; For the th multidimensional feature vector in the feature history buffer The mean of each dimension; For the th multidimensional feature vector in the feature history buffer Standard deviation of each dimension; The value is a very small positive number, used to prevent the denominator from being zero. In this embodiment of the invention, it is preset to... ; The mean Reflects the first under the current working conditions The normal baseline position of each dimension, the standard deviation Reflects the first The normal fluctuation range of the first dimension; the first dimension Deviation in each dimension The larger the value, the more significantly the current state deviates from the historical normal pattern; The deviation of each dimension in the multidimensional feature vector of the current sliding time window is concatenated to form the three-dimensional deviation vector of the current sliding time window; Furthermore, the above three-dimensional deviation is mapped to a comprehensive anomaly index. The purpose is that the deviation in each of the three dimensions reflects different aspects of the anomaly, but the subsequent threshold comparison requires a single scalar to determine whether the current sliding time window as a whole is within the normal range. Therefore, the three-dimensional information must be fused into a single value, and the fusion method uses Euclidean distance. In this embodiment of the invention, the square root of the sum of the squares of the deviations in all dimensions of the multi-dimensional feature vector of the current sliding time window is taken as the comprehensive anomaly index of the current sliding time window. The larger the comprehensive anomaly index, the further the current state deviates from the normal operating condition. It should be noted that, because the casting conditions drift slowly with factors such as ambient temperature and batch material differences, the comprehensive anomaly index under normal conditions is not a fixed value, but rather a statistical distribution that varies with the conditions. Therefore, a pre-calibrated fixed threshold cannot be used; the most recent historical data must be used to estimate the normal fluctuation level under the current conditions. However, the historical sliding time windows stored in the feature history buffer have already been judged as normal. Therefore, the mean and standard deviation of the comprehensive anomaly index of all historical sliding time windows in the feature history buffer are obtained to construct the dynamic discrimination threshold for the current sliding time window.

[0039] In this embodiment of the invention, the comprehensive anomaly index of each historical sliding time window in the feature history buffer is obtained; the mean and standard deviation of the comprehensive anomaly index of all historical sliding time windows in the feature history buffer are obtained. Obtain the dynamic discrimination threshold for the current sliding time window: ; in, The dynamic threshold for the current sliding time window; The average of the comprehensive anomaly index of all historical sliding time windows in the characteristic historical buffer; The standard deviation of the composite anomaly index of all historical sliding time windows in the characteristic historical buffer; The preset positive multiple coefficient is used. It should be noted that the discrimination threshold is called dynamic because the feature history buffer is constantly updated. When the working condition slowly drifts, the standard deviation and mean will also be updated accordingly. The dynamic discrimination threshold will also track the changes in the working condition, avoiding the problem of the fixed threshold becoming invalid after the working condition drifts. The The larger the value, the higher the threshold, the stricter the threshold for judging anomalies, the lower the false alarm rate but the higher the risk of missed alarms; The smaller the value, the lower the threshold and the higher the sensitivity to anomalies, but the false alarm rate increases. In this embodiment of the invention, a positive multiple coefficient is preset. In other embodiments, implementers may pre-determine specific implementation methods. The value; It should be noted that in the first few sliding time windows immediately after the casting begins, the feature history buffer has not yet been filled. There is a historical normal sliding time window. At this time, it is impossible to obtain the dynamic discrimination threshold of the sliding time window, so it is necessary to preset the initial threshold.

[0040] In this embodiment of the invention, if the historical normal sliding time window before the current sliding time window is not full... If there are [number], then the initial threshold for the current sliding time window is preset to [value]. The method for obtaining the initial threshold of the current sliding time window is as follows: within the process debugging cycle, obtain the comprehensive anomaly index corresponding to the initial sliding time window of no less than batch A of qualified casting batches, filter out the maximum comprehensive anomaly index value among all comprehensive anomaly indices, and multiply the maximum comprehensive anomaly index value by the safety factor as the initial threshold of the current sliding time window; the safety factor is 1.2. If the feature history buffer of the current sliding time window contains If the overall anomaly index of the current sliding time window exceeds the dynamic discrimination threshold of the current sliding time window, then the current sliding time window is determined to be an abnormal sliding time window, indicating the presence of preliminary abnormal signs, and an anomaly flag for the current sliding time window is output. Otherwise, output the exception flag for the current sliding time window. The current sliding time window is determined to be a normal sliding time window, which is within the normal fluctuation range; If the number of historical normal sliding time windows in the feature history buffer of the current sliding time window does not reach the required number... If the overall anomaly index of the current sliding time window is greater than the initial threshold of the current sliding time window, then the current sliding time window is determined to be an abnormal sliding time window, which shows preliminary abnormal signs, and the anomaly flag of the current sliding time window is output. Otherwise, output the exception flag for the current sliding time window. The current sliding time window is determined to be a normal sliding time window, which is within the normal fluctuation range.

[0041] S4: After confirming the validity of the abnormal state based on the determination results of multiple consecutive abnormal flags, generate corresponding graded process adjustment instructions according to the degree of deviation of the multidimensional feature vector from the normal baseline, and send them to the casting equipment actuator to intervene in the casting process in real time.

[0042] It should be noted that this step requires transforming the preliminary anomaly identification results into actionable adjustments to the casting process, achieving online defect suppression and closed-loop optimization of process parameters. In the preceding steps, through multi-sensor data preprocessing, multi-dimensional deviation feature extraction, and dynamic adaptive threshold comparison, early perception and quantitative assessment of abnormal states in the casting process have been achieved at the signal level. However, the preliminary anomaly identification results need to be transformed into actual control commands for the casting equipment, ensuring timely adjustment of key process parameters such as casting temperature, filling speed, and holding pressure during the defect incubation stage to truly achieve effective defect suppression. Therefore, this step uses the anomaly flags and their corresponding multi-dimensional feature vectors of each sliding time window as input, sequentially executing continuous confirmation and judgment of anomaly windows, anomaly level classification, and the generation and issuance of process parameter adjustment commands.

[0043] It should be further explained that, in order to avoid frequent malfunctions of the actuator due to the random characteristic jump of a single sliding time window, it is necessary to continuously confirm and judge the abnormal output flags.

[0044] In this embodiment of the invention, the preset number of consecutive confirmation windows is: If the current sliding time window is consecutive If the exception flag for each sliding time window is 1, the exception status of the current sliding time window is confirmed to be valid and subsequent actions are triggered; if the consecutive exception flags before the current sliding time window are 1, the exception status of the current sliding time window is confirmed to be valid and subsequent actions are triggered. If at least one anomaly flag is 0 within a sliding time window, the anomaly in the current sliding time window is determined to be an occasional disturbance, and no action is triggered. The system continues to maintain monitoring and waits for further verification in subsequent sliding time windows. In this embodiment of the invention, a preset number of consecutive confirmation windows is defined. The value is 10. In other implementations, the implementer can preset the value according to the specific implementation method. The value; The physical significance of the continuous confirmation mechanism lies in the fact that the development of casting defects is a continuous process driven by physical mechanisms such as impeded heat transfer and obstructed feeding sensing channels. This manifests as continuous or high-frequency appearance of abnormal indicators, rather than isolated single-point jumps. Therefore, continuous confirmation can effectively filter out transient interference caused by non-defect factors such as electrical noise and mechanical vibration, significantly reducing the false alarm rate and improving the system's reliability while ensuring that no real defects are missed. Furthermore, after confirming the validity of the abnormal state of the current sliding time window, the overall abnormal deviation degree of the current sliding time window is obtained: ; in, This represents the overall degree of abnormal deviation within the current sliding time window; The th feature vector in the current sliding time window The values ​​that can be taken in each dimension; For the th multidimensional feature vector in the feature history buffer The mean of each dimension; The value reflects the relative multiple by which the current value of each dimension deviates from its historical normal average. The larger the value, the more severe the deviation of that dimension. This anomaly intensity index... The magnitude of the value directly reflects the severity of the current abnormal state deviating from the boundary of normal operating conditions; Furthermore, the degree of comprehensive anomaly deviation is determined according to the preset anomaly level classification rules. Mapped to different anomaly levels; specifically, a first-level threshold is preset. Second-level threshold And satisfy ;when When, it is judged as a mild abnormality; when When, it is judged as moderately abnormal; when When this occurs, it is determined to be a severe abnormality; in this embodiment of the invention, a first-level threshold is preset. Second-level threshold In other embodiments, implementers may pre-set according to specific implementation conditions. as well as The value; Furthermore, based on the anomaly level determined above, corresponding process parameter adjustment instructions are generated.

[0045] For minor anomalies, a minor adjustment instruction is generated, which includes the following specific adjustment amount: adjusting the pouring temperature upwards within the preset allowable range. Adjust the filling speed downwards Adjust the holding pressure upwards. ,in, , , These represent the amounts of temperature increase, speed decrease, and pressure increase corresponding to minor adjustments. These adjustments aim to suppress the further development of defects by slightly correcting process parameters, while avoiding the introduction of new process fluctuations through large parameter changes. For moderate anomalies, a moderate adjustment command is generated, including adjusting the pouring temperature upwards. Adjust the filling speed downwards Adjust the holding pressure upwards. ,in, , , ;in, , , These correspond to the following adjustments for medium-amplitude adjustments: temperature increase, speed decrease, and pressure increase. For severe anomalies, a significant adjustment instruction is generated, including increasing the pouring temperature. Adjust the filling speed downwards Adjust the holding pressure upwards. ,in, , , , , , These adjustments correspond to significant increases in temperature, decreases in speed, and increases in pressure. Simultaneously, after these significant adjustments, the system continuously monitors the anomaly flags in subsequent windows. If the flags remain abnormal, a pouring termination command is triggered, initiating emergency pressure relief and pouring abortion operations.

[0046] The generated adjustment commands are sent to the temperature controller, flow control valve, and pressure regulating actuator of the casting equipment via standard industrial communication protocols, enabling real-time online intervention in the casting process.

[0047] It should be noted that the adjustment instructions involve , , , , , , , , The adjustment parameters can be preset according to the actual allowable range of the specific casting equipment and process window. It is necessary to ensure that each adjustment does not exceed the physical limits of the equipment's actuator and that the adjusted parameters remain within the effective process window. This embodiment of the invention does not impose excessive restrictions here; for example, for the low-pressure casting process of aluminum alloys, the following settings can be configured: 5℃ 10% of the current speed 5% of the current pressure 10℃ 20% of the current speed 10% of the current pressure 15℃ 30% of the current speed This represents 15% of the current pressure.

[0048] like Figure 2 As shown, the yield of the present invention is higher than that of the prior art.

[0049] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent suppression and parameter optimization of casting process defects, characterized in that, include: The original time-series data sequences of multiple sensor channels arranged at various measuring points of the mold during the casting process are obtained. The original time-series data sequences are standardized to obtain the standardized time-series data sequences of each sensor channel. The time-series data includes temperature, pressure and speed. The data is divided into a standardized time series by sliding a preset sliding time window. Statistical feature calculations are performed on the data of each sensor channel within each sliding time window. Based on the calculation results, the degree of deviation of each sensor channel signal from the normal operating condition is extracted, and the degree of deviation of each sensor channel is aggregated into a multi-dimensional feature vector of each sliding time window. The multidimensional feature vectors of each sliding time window are sequentially input into the judgment model. The judgment model uses the multidimensional feature vectors of multiple sliding time windows that were judged to be in a normal state before the current sliding time window to form a feature history buffer. The dynamic discrimination threshold is estimated based on the distribution of feature vectors in the feature history buffer. The deviation of the multidimensional feature vector of the current sliding time window from the dynamic discrimination threshold is used as the basis for the anomaly judgment of the current sliding time window, and an anomaly flag is output. The judgment model is constructed as follows: a feature history buffer is formed by constructing a multi-dimensional feature vector of multiple consecutive sliding time windows that were judged to be in a normal state before the current sliding time window; for each dimension of the multi-dimensional feature vector of the current sliding time window, the deviation of the value of that dimension relative to the mean and standard deviation of the corresponding dimension in the feature history buffer is calculated, and the deviations of each dimension are concatenated into a deviation vector; the square root of the sum of squares of each dimension of the deviation vector is taken to obtain the comprehensive anomaly index of the current sliding time window; the mean and standard deviation of the comprehensive anomaly index of all historical sliding time windows in the feature history buffer are calculated, and the sum of the preset multiples of the mean and standard deviation is used as the dynamic judgment threshold of the current sliding time window. After confirming the validity of the abnormal state based on the judgment results of multiple consecutive abnormal indicators, corresponding graded process adjustment instructions are generated according to the degree of deviation of the multidimensional feature vector from the normal baseline.

2. The intelligent suppression and parameter optimization method for casting process defects according to claim 1, characterized in that, The standardized time-series data sequences of each sensing channel are obtained in the following manner: The original sequence of each sensing channel is obtained by collecting data at a preset sampling frequency until the end of the pouring process; any sensing channel is selected as a reference sensing channel, and the time delay of each sensing channel relative to the reference sensing channel is obtained by cross-correlation time delay estimation. The original time series data sequence of each sensing channel is time-shifted and compensated according to the time delay to eliminate the time deviation between each sensing channel. After time-shift compensation, the data from each sensor channel are smoothed and denoised using a sliding filter window to obtain the filtered time-series data sequence for each sensor channel. The global mean and standard deviation of the filtered time-series data sequence for each sensor channel are calculated. The data value at each sampling time is subtracted from the global mean and then divided by the standard deviation to obtain the standardized time-series data sequence for each sensor channel.

3. The intelligent suppression and parameter optimization method for casting process defects according to claim 1, characterized in that, The multidimensional feature vector is obtained in the following manner: The standardized time-series data is divided by a preset sliding time window. For each sliding time window, the mean and standard deviation of the data from each sensing channel within the window, as well as the linear trend coefficient of the fitted line, are calculated. The instantaneous deviation is obtained by the deviation of the last time value within the window from the mean within the window; the trend deviation is obtained by the difference between the linear trend coefficient and the historical normal reference trend coefficient; and the fluctuation distortion is obtained by the relative deviation of the standard deviation from the historical normal reference standard deviation. The three deviations of each channel within the same window are arranged into three vectors, and the norm of each vector is calculated as an anomaly intensity scalar for each dimension and concatenated into a multidimensional feature vector.

4. The intelligent suppression and parameter optimization method for casting process defects according to claim 3, characterized in that, The methods for obtaining the historical normal reference trend coefficient and the historical normal reference standard deviation are as follows: A batch parameter A is preset, and at least A batches of qualified casting batches from continuous production that have been confirmed by non-destructive testing to be free of internal defects are selected. For each qualified batch, obtain the standardized time-series data sequence of each sensor channel throughout the entire casting process, and calculate the linear trend coefficient and data standard deviation of each sensor channel throughout the entire casting process. The mean of the linear trend coefficients of the same sensor channel across all qualified batches is taken as the historical normal reference trend coefficient of that sensor channel, and the mean of the standard deviations of the data of the same sensor channel across all qualified batches is taken as the historical normal reference standard deviation of that sensor channel.

5. The intelligent suppression and parameter optimization method for casting process defects according to claim 1, characterized in that, The feature history buffer is updated as follows: if the current sliding time window is determined to be normal, its multidimensional feature vector is stored in the buffer and the earliest stored feature vector is removed; if the current sliding time window is determined to be abnormal, the buffer is not updated.

6. The intelligent suppression and parameter optimization method for casting process defects according to claim 1, characterized in that, When the number of historical sliding time windows stored in the feature history buffer does not reach the preset capacity, a preset initial threshold is used as the discrimination threshold for the current sliding time window. The initial threshold is obtained by taking the product of the maximum value of the comprehensive anomaly index corresponding to the initial sliding time window of no less than batch A of qualified casting batches within the process debugging cycle and the safety factor.

7. The intelligent suppression and parameter optimization method for casting process defects according to claim 1, characterized in that, The step of confirming the validity of the abnormal state based on the determination results of multiple consecutive abnormal flags includes: The number of consecutive confirmation windows is preset. If the abnormal flags of the previous consecutive confirmation windows are all state values ​​that indicate an abnormality, the abnormal state of the current sliding window is confirmed to be valid. If at least one abnormal flag representing a normal state value exists in several consecutive confirmation windows preceding the current sliding time window, then the current abnormality is determined to be an occasional disturbance, and no action is triggered.

8. The intelligent suppression and parameter optimization method for casting process defects according to claim 1, characterized in that, The generation of the corresponding graded process adjustment instruction includes: The sum of the relative deviations of the values ​​of each dimension in the multidimensional feature vector of the current sliding time window from the mean of each dimension in the feature history buffer is used as the comprehensive degree of abnormal deviation. A first-level threshold and a second-level threshold are preset. When the deviation does not exceed the first-level threshold, a micro-adjustment instruction containing a first adjustment amount is generated. When the deviation exceeds the first-level threshold but does not exceed the second-level threshold, a medium-level adjustment instruction containing a second adjustment amount is generated. When the deviation exceeds the second-level threshold, a large-scale adjustment instruction containing a third adjustment amount is generated. The second adjustment amount is greater than the first adjustment amount, the third adjustment amount is greater than the second adjustment amount, and the large-scale adjustment instruction also includes triggering a pouring termination instruction if the continuous judgment after execution is still abnormal.

9. A smart system for suppressing defects and optimizing parameters in the casting process, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the intelligent suppression and parameter optimization method for casting process defects according to any one of claims 1-8.

Images