Intelligent evaluation method for pouring mold life based on multi-source data analysis

Abstract

The application discloses a pouring mold life intelligent evaluation method based on multi-source data analysis, and relates to the technical field of data analysis. The method comprises the following steps: synchronously collecting multi-dimensional state data of a pouring mold in real time, obtaining a multi-source heterogeneous state data set for time sequence alignment, and generating an aligned data set; decoupling multi-scale degradation features based on the aligned data set, and constructing a multi-modal degradation feature vector; constructing an adaptive graph attention fusion network to perform heterogeneous feature fusion on the multi-modal degradation feature vector, and generating a mold comprehensive health state index curve; predicting the remaining service life of the pouring mold based on the comprehensive health state index curve, correlating the mold remaining life evaluation result with pouring process parameters, and generating a mold life evaluation report. The technical problem that the life change trend of a pouring mold is difficult to accurately evaluate in the prior art during long-term pouring operation is solved, and the technical effect of improving the reliability of pouring mold life evaluation is achieved.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and more specifically to an intelligent assessment method for the lifespan of casting molds based on multi-source data analysis. Background Technology

[0002] As a critical load-bearing component in the casting production process, casting molds are prone to wear, fatigue cracking, thermal deformation, and localized structural degradation under long-term high temperature, high pressure, and cyclic thermal shock environments. Their operational status directly affects the casting quality and production stability. With the increase in the number of consecutive castings, the degradation characteristics inside the mold typically exhibit non-linear cumulative changes. Relying solely on manual inspections or single-parameter monitoring methods is insufficient to accurately reflect the true health status of the mold. Existing casting mold life assessment technologies mostly employ fixed threshold alarms, empirical life estimations, or single-dimensional sensor data analysis for status judgment. These methods suffer from insufficient utilization of multi-source heterogeneous data such as temperature, vibration, pressure, and operating process parameters, making it difficult to effectively reveal the correlation and degradation patterns between different state parameters. Furthermore, under complex operating conditions, the mold degradation process exhibits significant temporal coupling and dynamic fluctuation characteristics. Traditional assessment methods lack the ability to comprehensively analyze multi-scale degradation characteristics, easily leading to large life assessment biases and insufficient reliability in remaining service life predictions, thus affecting the accuracy of casting mold maintenance decisions and the stability of the production process. Summary of the Invention

[0003] This application provides an intelligent assessment method for the lifespan of casting molds based on multi-source data analysis, which solves the technical problem that it is difficult to accurately assess the lifespan change trend of casting molds during long-term casting operation in the prior art.

[0004] This application provides a method for intelligent assessment of the lifespan of casting molds based on multi-source data analysis, the method comprising:

[0005] Multi-dimensional state data of the casting mold are collected in real time and synchronously to obtain a multi-source heterogeneous state dataset. This dataset is then time-series aligned to generate an aligned dataset. Based on the aligned dataset, multi-scale degradation features are decoupled to construct a multi-modal degradation feature vector. An adaptive graph attention fusion network is constructed to fuse the heterogeneous features of the multi-modal degradation feature vector, generating a comprehensive health status index curve for the mold. Based on the comprehensive health status index curve, the remaining service life of the casting mold is predicted. The remaining service life assessment results are then correlated with the casting process parameters to generate a mold service life assessment report.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, multi-dimensional state data of the casting mold is collected in real time and synchronously to obtain a multi-source heterogeneous state dataset. This dataset is then time-series aligned to generate an aligned dataset. Next, multi-scale degradation features are decoupled based on the aligned dataset to construct a multi-modal degradation feature vector. Then, an adaptive graph attention fusion network is constructed to fuse heterogeneous features of the multi-modal degradation feature vector, generating a comprehensive health status index curve for the mold. Finally, the remaining service life of the casting mold is predicted based on the comprehensive health status index curve. The remaining service life assessment results are correlated with casting process parameters to generate a mold service life assessment report. This solves the technical problem of accurately assessing the service life trend of casting molds during long-term casting operation in existing technologies, achieving the technical effect of improving the reliability of casting mold service life assessment. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 A schematic diagram of the intelligent assessment method for casting mold life based on multi-source data analysis provided in this application embodiment; Figure 2 This is a schematic diagram of the process for generating the comprehensive health status index curve of the casting mold in the intelligent assessment method for casting mold life based on multi-source data analysis provided in the embodiments of this application. Detailed Implementation

[0009] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0010] Examples, such as Figure 1 As shown, this application provides an intelligent assessment method for the lifespan of casting molds based on multi-source data analysis, wherein the method includes: Real-time synchronous acquisition of multi-dimensional state data of casting molds, obtaining multi-source heterogeneous state datasets, performing time-series alignment, and generating aligned datasets.

[0011] Furthermore, the method involves real-time synchronous acquisition of multi-dimensional state data from the casting mold, obtaining a multi-source heterogeneous state dataset, performing time-series alignment, and generating an aligned dataset. This includes: Based on a multi-source sensor network configured on the casting production line, the system acquires mold surface identification images through the multi-source sensor network; performs identity resolution based on the mold surface identification images to obtain mold identity codes; performs synchronous monitoring according to the mold identity codes to obtain mold closing trigger signals; and uses the mold closing trigger signals as a global synchronization reference to synchronously collect and broadcast to the multi-source sensor network, driving the multi-source sensor network to perform trigger sampling and generate the multi-source heterogeneous state dataset.

[0012] A multi-source sensor network is configured for each casting station on the casting production line. The multi-source sensor network includes at least temperature sensors, vibration sensors, pressure sensors, acoustic emission sensors, current sensors, and an industrial vision acquisition device. The temperature sensors are used to collect temperature data of the mold cavity area and cooling area; the vibration sensors are used to collect vibration data during the mold opening and closing process; the pressure sensors are used to collect mold closing pressure and casting pressure data; the acoustic emission sensors are used to collect acoustic emission data generated by the propagation of cracks inside the mold; the current sensors are used to collect operating status data of the drive mechanism; and the industrial vision acquisition device is used to collect wear images and crack images on the mold surface. A unique identification mark is pre-set on the surface of each casting mold. The identification mark is read by industrial vision recognition or radio frequency identification to obtain the corresponding mold identification code.

[0013] After obtaining the mold identification code, the operating status of the corresponding casting equipment is synchronously monitored according to the mold identification code, and the mold closing displacement signal, hydraulic pressure signal, and limit switch status signal in the PLC control system are read in real time. When the mold closing displacement reaches the preset closing threshold and the hydraulic pressure reaches the preset mold closing pressure threshold, it is determined that the current mold has entered the effective mold closing state, and a corresponding mold closing trigger signal is generated. The mold closing trigger signal is used as a unified time reference for synchronous acquisition and broadcast to each sensor acquisition device. After receiving the mold closing trigger signal, each sensor acquisition device synchronously starts to acquire the corresponding status parameters. Among them, the temperature sensor is used to acquire mold cavity temperature data, the vibration sensor is used to acquire mold vibration spectrum data, the pressure sensor is used to acquire mold closing pressure and casting pressure data, the acoustic emission sensor is used to acquire internal impact and crack propagation signals of the mold, and the current acquisition device is used to acquire the driving mechanism operating current data. During the sampling process, each sensor acquisition device adds a sampling timestamp according to a unified synchronous clock, and the sampling results are centrally cached and uniformly integrated to generate a multi-source heterogeneous status dataset with a unified time reference.

[0014] Furthermore, methods for obtaining multi-source heterogeneous state datasets and performing time-series alignment to generate aligned datasets include: Based on the multi-source heterogeneous state dataset, command response delay analysis is performed to generate command response delay parameters; the multi-source heterogeneous state dataset is time-corrected according to the command response delay parameters to determine multiple data timestamps; the multi-source heterogeneous state dataset is segmented according to process stage windows to generate multiple process segment data; the multiple process segment data are linearly time-aligned according to the multiple data timestamps to generate the aligned dataset.

[0015] Preferably, the original sampling time corresponding to various state data in the multi-source heterogeneous state dataset is obtained, and the first response time of each sensor acquisition device after receiving the mold closing trigger signal is extracted. Response delay analysis is performed based on the time difference between the first response time and the transmission time of the mold closing trigger signal to calculate the acquisition response delay value corresponding to each type of state data. Combined with the data transmission time, cache write time, and sampling start time of different sensor acquisition devices, the acquisition response delay value is cumulatively compensated to generate corresponding command response delay parameters. The original sampling time corresponding to each type of state data is time-corrected according to the command response delay parameters, and the corrected sampling time is used as the data timestamp of the corresponding state data, thereby determining the data timestamp under multiple unified time bases. Subsequently, the multi-source heterogeneous state dataset is divided into process stages according to the casting process flow. The process stages include at least a preheating stage, a mold-closing stage, a casting stage, a pressure-holding stage, a cooling stage, and a demolding stage. Multiple process stage windows are constructed according to the start and end times of each stage. Based on these multiple process stage windows, the multi-source heterogeneous state dataset is segmented, dividing state data belonging to the same process stage window into corresponding process segment data. For process segment data obtained at different sampling frequencies, a unified time axis is established according to the multiple data timestamps, and linear interpolation is used to fill in missing time points, ensuring that temperature data, vibration data, pressure data, acoustic emission data, and current data from different sources correspond to each other on a unified time coordinate. Linear temporal alignment is performed on each process segment data according to the unified time axis, and the data is then spliced ​​and integrated in chronological order to finally generate an aligned dataset.

[0016] Based on the aligned dataset, multi-scale degradation features are decoupled to construct multi-modal degradation feature vectors.

[0017] Furthermore, based on the aligned dataset, multi-scale degradation feature decoupling is performed to construct a multi-modal degradation feature vector. The method includes: The aligned dataset is traversed to perform parameter adaptive variational mode decomposition, generating a multi-scale intrinsic mode function component set; the multi-scale intrinsic mode function component set is traversed to perform degradation sensitivity assessment, calculating a degradation sensitivity score; based on the degradation sensitivity score, the aligned dataset is traversed to perform modality degradation matching, extracting multiple sensitive components; the multiple sensitive components are divided into two-stage scales according to instantaneous frequency, generating a high-frequency degradation component set and a low-frequency degradation component set; based on the high-frequency degradation component set, cyclic feature capture is performed to extract local degradation features; based on the low-frequency degradation component set, trend feature capture is performed to extract global degradation features; the local degradation features and the global degradation features are concatenated across modes to generate the multi-modal degradation feature vector.

[0018] Preferably, temperature data, vibration data, pressure data, acoustic emission data, and driving current data from the aligned dataset are acquired, and corresponding state time-series signals are constructed according to a unified time axis. A sliding time window is used to traverse each state time-series signal, and a fast Fourier transform is performed on the traversed state time-series signal to obtain the corresponding spectral distribution results. Variational mode decomposition parameters are adaptively determined based on the main energy concentration region of the spectrum, the frequency dispersion, and the signal fluctuation amplitude. The frequency dispersion is calculated using the spectral energy variance, and the signal fluctuation amplitude is calculated using the difference between the maximum and minimum values ​​within the time window. After determining the number of mode decomposition layers, the penalty factor, and the iteration convergence threshold, parameter-adaptive variational mode decomposition is performed on each state time-series signal, decomposing the original state time-series signal into multiple intrinsic mode function components corresponding to different frequency scales, and generating a multi-scale intrinsic mode function component set.

[0019] A degradation sensitivity assessment is performed on each intrinsic mode function (EMF) component in the multi-scale intrinsic mode function component set. Short-time energy volatility, frequency drift rate, and periodic stability coefficient are calculated. The short-time energy volatility is calculated as the ratio of the energy difference between adjacent time windows to the average energy of the previous time window. The frequency drift rate is calculated as the ratio of the difference between the current center frequency and the initial center frequency to the initial center frequency. The periodic stability coefficient is calculated as the ratio of the periodic standard deviation to the periodic mean. A reference degradation feature sequence is retrieved from historical mold failure samples, and the correlation matching degree between the current EMF component and the reference degradation feature sequence is calculated using the Pearson correlation coefficient. The short-time energy volatility, frequency drift rate, periodic stability coefficient, and correlation matching degree are weighted and fused according to preset weights to obtain the corresponding degradation sensitivity score. EMF components with degradation sensitivity scores higher than a preset sensitivity threshold are extracted as degradation sensitive components.

[0020] Perform a Hilbert transform on the degradation-sensitive component to extract the corresponding instantaneous frequency sequence; obtain the reference frequency corresponding to the mold opening and closing cycle, and calculate the median frequency of the instantaneous frequency sequence corresponding to each degradation-sensitive component; when the median frequency is higher than the reference frequency, the corresponding degradation-sensitive component is assigned to the high-frequency degradation component set; when the median frequency is lower than or equal to the reference frequency, the corresponding degradation-sensitive component is assigned to the low-frequency degradation component set.

[0021] Local cyclic degradation features are extracted for the set of high-frequency degradation components. The local impact peak is obtained by statistically analyzing the instantaneous maximum amplitude of the high-frequency degradation components. The short-time energy mutation rate is calculated by the energy change rate of adjacent sampling windows. The spectral sideband expansion width is calculated by the difference between the frequency ranges of the sidebands on both sides of the main frequency. The cyclic impact period is calculated by the average time interval between adjacent impact peaks. Local degradation features are generated based on the local impact peak, short-time energy mutation rate, spectral sideband expansion width, and cyclic impact period.

[0022] Long-term trend degradation features are extracted for the set of low-frequency degradation components. Local weighted regression fitting is performed on each low-frequency degradation component to obtain the corresponding degradation trend curve. The average monotonic slope is calculated by averaging the slopes of adjacent sampling points in the degradation trend curve, the cumulative attenuation amplitude is calculated by the difference between the initial value and the current value of the degradation trend curve, and the period drift is calculated by the offset between the current main period and the initial main period. Global degradation features are generated based on the average monotonic slope, cumulative attenuation amplitude, and period drift.

[0023] The local degradation features and the global degradation features are normalized according to a unified feature dimension, and cross-modal splicing and integration are performed according to a unified time order to generate a multimodal degradation feature vector for mold health status assessment and remaining life prediction.

[0024] Furthermore, the method involves dividing the multiple sensitive components into a two-stage scale based on instantaneous frequency to generate a set of high-frequency degradation components and a set of low-frequency degradation components. Based on the multiple sensitive components, a Hilbert transform is performed to construct an instantaneous frequency sequence; the instantaneous frequency sequence is traversed to identify the median and determine the instantaneous frequency median; a mold opening and closing cycle is introduced, which includes a reference frequency, and sensitive components are extracted based on the reference frequency; the instantaneous frequency median is compared according to the mold opening and closing cycle: S1: the instantaneous frequency median higher than the sensitive component is assigned to the high-frequency degradation component set; S2: the instantaneous frequency median lower than or equal to the sensitive component is assigned to the low-frequency degradation component set.

[0025] Preferably, the multiple sensitive components are subjected to Hilbert transforms respectively, and the time-domain signals corresponding to each sensitive component are convolved to obtain the corresponding orthogonal imaginary part signals. The original real part signals are then combined with the orthogonal imaginary part signals to generate an analytical signal. Based on the analytical signal, the corresponding phase angle sequence is calculated, wherein the phase angle sequence is obtained by solving the imaginary and real part signals using the arctangent function. Subsequently, the phase angle sequence is differentially processed according to the sampling time interval, and the difference result is multiplied by the sampling frequency and divided by 2π to obtain the instantaneous frequency value at the corresponding moment, thereby constructing the instantaneous frequency sequence of the corresponding sensitive component. A sliding time window is traversed for each instantaneous frequency value in the instantaneous frequency sequence, and the average frequency value within the current sliding time window is calculated. When any instantaneous frequency value deviates from the average frequency value of the current sliding time window by more than a preset offset threshold, the corresponding instantaneous frequency value is determined as an abnormal frequency point and filtered out. Subsequently, the instantaneous frequency values ​​after filtering out abnormal frequency points are sorted in ascending order of numerical value, and the result is extracted. The frequency value at the middle position after the column is taken as the median of the instantaneous frequency of the corresponding sensitive component. The start time of mold closing and the end time of mold opening of the casting equipment in multiple consecutive production cycles are obtained, and the time interval between two adjacent mold closing actions is calculated. The average of multiple time intervals is used to obtain the mold opening and closing cycle time, and the corresponding reference frequency is calculated based on 1 divided by the mold opening and closing cycle time. Then, the median of the instantaneous frequency corresponding to each sensitive component is compared with the reference frequency one by one. When the median of the instantaneous frequency is greater than the reference frequency, the corresponding sensitive component is written into the high-frequency degradation component buffer area, and the corresponding high-frequency degradation component index is established. When the median of the instantaneous frequency is less than or equal to the reference frequency, the corresponding sensitive component is written into the low-frequency degradation component buffer area, and the corresponding low-frequency degradation component index is established. After completing the traversal of all sensitive components, all sensitive components in the high-frequency degradation component buffer area are integrated to generate a high-frequency degradation component set, and all sensitive components in the low-frequency degradation component buffer area are integrated to generate a low-frequency degradation component set.

[0026] Furthermore, based on the set of low-frequency degradation components, trend feature capture is performed to extract global degradation features. The method includes: A degradation trend curve is constructed by performing local weighted regression based on the set of low-frequency degradation components; the average monotonic slope of the degradation trend curve is calculated based on the degradation trend curve; a descent analysis is performed by mapping the average monotonic slope to the degradation trend curve to extract the cumulative descent magnitude; the degradation trend curve is combined with the cumulative descent magnitude for attenuation fitting to extract the fitting structural parameters; and the fitting structural parameters are added to the global degradation feature.

[0027] Preferably, the low-frequency degradation components in the low-frequency degradation component set are arranged in chronological order, and a sliding time window is used to perform local weighted regression on each low-frequency degradation component. Data points within the neighborhood of the current time point are used as local fitting samples, and corresponding weights are assigned based on the time distance between each data point and the current time point; the smaller the time distance, the higher the corresponding weight. Subsequently, the local fitting samples are regressed using the weighted least squares method to obtain the local fitting results corresponding to each time position. These local fitting results are then concatenated in chronological order to construct a degradation trend curve. The slope of change between adjacent sampling points in the degradation trend curve is calculated, and all slopes are averaged to obtain the average monotonic slope of the degradation trend curve. The slope of change between adjacent sampling points is determined by the sum of the values ​​of the current sampling point and... The difference between the values ​​of the previous sampling points is divided by the corresponding time difference to obtain the value. Then, the average monotonic slope is mapped to the degradation trend curve. The descending segments of the degradation trend curve that are lower than the trend line corresponding to the average monotonic slope are traversed and statistically analyzed. The amplitude difference between the start and end points of each descending segment is calculated, and all amplitude differences are accumulated to obtain the cumulative decline amplitude. Furthermore, the degradation trend curve and the cumulative decline amplitude are synchronously input into the decay fitting model to perform exponential decay fitting on the degradation trend curve. The fitting structure parameters corresponding to the minimum error between the fitted curve and the degradation trend curve are solved by iterative minimum error method. The fitting structure parameters include at least the decay coefficient, decay rate, and trend stability coefficient. Finally, the fitting structure parameters are written into the global degradation feature vector according to the preset feature dimensions to generate the corresponding global degradation features.

[0028] An adaptive graph attention fusion network is constructed to perform heterogeneous feature fusion on the multimodal degradation feature vectors, generating a comprehensive health status index curve for the mold.

[0029] Furthermore, such as Figure 2 As shown, an adaptive graph attention fusion network is constructed to perform heterogeneous feature fusion on the multimodal degradation feature vectors to generate a comprehensive health status index curve for the mold. The method includes: The local degradation features and the global degradation features are each treated as independent feature nodes to form a feature node set. The feature node set is traversed to calculate node mutual information and node correlation coefficient. Based on the node mutual information and node correlation coefficient, a weighted combination is performed to construct the node connection strength value. The node connection strength value is mapped to the feature node set for association matching to construct an initial fully connected feature relationship graph. The initial fully connected feature relationship graph is combined with the multimodal degradation feature vector and synchronized to the adaptive graph attention fusion network for heterogeneous feature fusion to generate the mold comprehensive health status index curve.

[0030] Preferably, the local impact peak, short-time energy mutation rate, spectral sideband expansion width, and cyclic impact period in the local degradation features are respectively used as local feature nodes, and the average monotonic slope, cumulative attenuation amplitude, periodic drift, and attenuation coefficient in the global degradation features are respectively used as global feature nodes, thus forming a feature node set; then, any two feature nodes in the feature node set are combined and traversed, and the node mutual information is calculated based on the historical change sequence of the corresponding feature nodes, wherein the node mutual information is calculated by the log ratio between the joint probability distribution and the marginal probability distribution, and is used to characterize the degree of information dependence between the two feature nodes; further based on the Peel method... The correlation coefficient is calculated to determine the node correlation coefficient between corresponding feature nodes. The node correlation coefficient is calculated as the ratio between the product of the covariance and the standard deviation of the change sequences of two feature nodes, and is used to characterize the degree of linear correlation between the two feature nodes. Then, the node mutual information and the node correlation coefficient are weighted and fused according to preset weights to obtain the node connection strength value between corresponding feature nodes, and the node connection strength value is written into the corresponding adjacency matrix position. After completing the traversal of all feature nodes, the connection relationship between feature nodes is established according to the adjacency matrix, and the connection strength values ​​between each feature node and the corresponding node are associated and mapped to construct an initial fully connected feature relationship graph.

[0031] Furthermore, the initial fully connected feature graph is combined with the multimodal degradation feature vector and synchronized to the adaptive graph attention fusion network for heterogeneous feature fusion to generate the comprehensive health status index curve of the mold. The method includes: An adaptive graph attention fusion network is constructed, comprising a multi-head graph attention layer, a temporal convolutional layer, and a feature aggregation layer. The multimodal degradation features are located according to the initial fully connected features to determine real-time degradation stage features. The multi-head graph attention layer dynamically adjusts the weights of multiple attention heads according to the real-time degradation node features to generate stage-adaptive fusion features. The temporal convolutional layer performs temporal smoothing on the stage-adaptive fusion features to generate multiple node temporal features. The feature aggregation layer compresses and aggregates the multiple node temporal features to construct a one-dimensional degradation feature sequence. A degradation monotonicity constraint is applied to the one-dimensional degradation feature sequence, and regression mapping is performed based on the constraint results to extract a monotonically decreasing curve as the comprehensive health status index curve of the model.

[0032] Preferably, an adaptive graph attention fusion network is constructed, wherein the number of input nodes is determined based on the number of feature dimensions of the multimodal degenerate feature vectors, and a corresponding graph structure input layer is established according to the number of input nodes; graph connection relationships are constructed based on the adjacency matrix corresponding to the connection strength values ​​between nodes, and the connection strength values ​​in the adjacency matrix are written into the corresponding graph edge weights; subsequently, a multi-head graph attention layer, a temporal convolutional layer, and a feature aggregation layer are sequentially built after the graph structure input layer, wherein each attention head in the multi-head graph attention layer corresponds to an independent node weight matrix and attention parameter matrix, and the temporal convolutional layer adopts a one-dimensional convolutional structure, and the convolutional window length is set according to the mold opening and closing cycle. The feature aggregation layer adopts a hybrid aggregation structure of global average pooling and max pooling. Furthermore, the adaptive graph attention fusion network is trained using historical mold degradation samples for parameter initialization. Specifically, the multimodal degradation feature vectors corresponding to the historical mold degradation samples are input into the graph structure input layer, and the actual lifetime degradation curve of the historical mold is used as the supervision label. The node weight matrix, attention parameter matrix, and convolution kernel parameters in the temporal convolutional layer are iteratively updated through backpropagation. When the change in the loss function between two consecutive training iterations is lower than a preset convergence threshold, parameter iteration is stopped, and the construction of the adaptive graph attention fusion network is completed.

[0033] Preferably, the node connection relationships in the initial fully connected feature graph are used as the graph structure input, and the multimodal degradation feature vectors are used as the input feature vectors of the corresponding nodes, which are simultaneously input into the adaptive graph attention fusion network. Then, based on the connection strength values ​​between nodes in the initial fully connected feature graph, a neighborhood traversal is performed on each feature node, and neighboring nodes with connection strength values ​​higher than a preset connection threshold are extracted as associated nodes of the current node, thereby determining the corresponding real-time degradation stage features. Further, the current node features and associated node features are input into multiple independent attention heads, a linear transformation is performed on the node features in each attention head, and the transformed node features are concatenated. This is then processed by Leakage. The yReLU activation function is used to calculate the attention score between nodes. Then, the attention score between nodes is weighted and corrected by combining the connection strength value between nodes, and normalized by the Softmax function to obtain the node attention weight under the corresponding attention head. The features of the associated nodes are weighted and summed according to the attention weight of each node to obtain the feature aggregation result under the corresponding attention head. The feature fluctuation amplitude and local degradation change rate in the feature aggregation result of each attention head are further statistically analyzed, and the output weight of each attention head is dynamically adjusted according to the statistical results. Then, the feature aggregation results of multiple attention heads are weighted and fused according to the adjusted output weight to generate stage adaptive fusion features.

[0034] The stage-adaptive fusion features are input into a temporal convolutional layer, and a one-dimensional convolutional kernel is used to perform sliding convolution along the time axis to continuously convolve the fusion features in adjacent time windows. Then, nonlinear activation and temporal smoothing are applied to the convolutional output to generate multiple node-specific temporal features. These multiple node-specific temporal features are further input into a feature aggregation layer, where global average pooling and max pooling are performed on each node-specific temporal feature. The pooling results are then concatenated and compressed to construct a corresponding one-dimensional degradation feature sequence. This one-dimensional degradation feature sequence is traversed in chronological order, and the difference in change between adjacent time points is calculated. When the feature value corresponding to a later time point is greater than the feature value corresponding to a previous time point, a monotonically decreasing penalty value is applied to the corresponding time point, and a degradation monotonicity constraint function is constructed based on all penalty values. Subsequently, an order-preserving regression mapping is performed on the one-dimensional degradation feature sequence based on the degradation monotonicity constraint function, and the regression results are calculated for error convergence through iterative optimization. The regression curve that satisfies the overall monotonically decreasing trend is extracted as the mold comprehensive health status index curve.

[0035] Furthermore, regression mapping is performed based on the constraint results to extract the monotonically decreasing curve. Methods include: Based on the constraint results, an order-preserving regression is performed on the zero-to-one interval to construct a nonparametric mapping function; inference is then performed according to the nonparametric mapping function to extract the monotonically decreasing curve.

[0036] Preferably, the constraint results corresponding to the degenerate monotonicity constraint function are obtained, and each feature value in the one-dimensional degenerate feature sequence is normalized. Specifically, the minimum feature value in the feature sequence is subtracted from the current feature value, and then divided by the difference between the maximum and minimum feature values, thus mapping all feature values ​​to the interval between zero and one. Subsequently, the mapped feature values ​​are arranged in chronological order, and order-preserving regression is performed based on the arranged feature values. Specifically, the regression value corresponding to each time position is initialized to the mapped feature value corresponding to the current time position, and the monotonicity of the regression values ​​corresponding to adjacent time positions is judged. When the regression value corresponding to the later time position is greater than the regression value corresponding to the previous time position, interval merging is performed on the corresponding time segment, and all regression values ​​in the merged interval are summed. The weighted average is updated; then, the monotonicity judgment and interval merging process of the regression values ​​at adjacent time positions are repeated until the regression values ​​corresponding to all time positions satisfy the monotonically decreasing constraint. Further, a nonparametric mapping function is constructed based on the regression values ​​that satisfy the monotonically decreasing constraint, where the time position is used as the input variable and the corresponding regression value is used as the output variable. A piecewise linear interpolation method is used to establish a continuous mapping relationship between each time interval. Finally, the one-dimensional degradation feature sequence is input into the nonparametric mapping function in chronological order, and regression inference is performed on the degradation feature values ​​corresponding to each time position to obtain a continuous regression result that satisfies the overall monotonically decreasing constraint. The continuous regression results are then connected in chronological order, and the corresponding monotonically decreasing curve is extracted as the mold comprehensive health status index curve.

[0037] Based on the comprehensive health status index curve, the remaining service life of the casting mold is predicted, and the remaining service life assessment results of the mold are correlated with the casting process parameters to generate a mold service life assessment report.

[0038] Furthermore, based on the comprehensive health status index curve, the remaining service life of the casting mold is predicted to obtain the mold's remaining service life assessment result. The method includes: The comprehensive health status index curve is subjected to hybrid degradation trajectory prediction to generate a degradation trajectory prediction sequence; based on the degradation trajectory prediction sequence, time domain positioning is performed to determine the failure prediction time value; the real-time time value and the failure prediction time value are retrieved and the time difference is calculated to determine the remaining usable mold times data; based on the remaining usable mold times data, synchronous quantization is performed to generate the mold remaining life assessment result.

[0039] Preferably, the comprehensive health status index curve of the mold is obtained, and the health status index values ​​corresponding to each sampling time are extracted in chronological order to construct a health status time series. Then, a sliding time window is used to segment and traverse the health status time series, and the degradation slope, local fluctuation amplitude, and trend decay rate are calculated for each time window. The degradation slope is calculated as the ratio between the difference between the first and last health status indices in the current time window and the corresponding time difference; the local fluctuation amplitude is calculated as the difference between the maximum and minimum health status index values ​​in the current time window; and the trend decay rate is calculated as the change in the average degradation slope between adjacent time windows. Subsequently, the degradation slope, local fluctuation amplitude, and trend decay rate are synchronously input into a hybrid degradation trajectory prediction model. This model includes a long-term trend prediction branch and a short-term fluctuation prediction branch. The long-term trend prediction branch is used to predict the overall degradation trend of the health status index, and the short-term fluctuation prediction branch is used to predict the local fluctuation changes of the health status index. Furthermore, the long-term trend prediction results and the short-term fluctuation prediction results are time-aligned and weighted to generate a corresponding degradation trajectory prediction sequence.

[0040] After obtaining the degradation trajectory prediction sequence, the degradation trajectory prediction sequence is traversed in chronological order, and a corresponding mold failure threshold is preset. Then, it is determined whether the health status index corresponding to each prediction time is lower than the mold failure threshold. When the predicted health status index in the degradation trajectory prediction sequence is lower than the mold failure threshold for the first time, the corresponding time point is determined as the failure prediction time value. The current real-time time value is further obtained, and the time difference between the failure prediction time value and the current real-time time value is calculated to obtain the remaining available time of the mold. Based on this, the current casting production cycle is further obtained, and the remaining available time of the mold is converted into the corresponding remaining available mold number data according to the average number of production cycles per unit time.

[0041] After obtaining the remaining usable number of mold cycles, the remaining usable number of mold cycles is subjected to interval quantization processing. Multiple life assessment level intervals are preset, corresponding to normal use status, early warning maintenance status, and critical failure status, respectively. Then, the life assessment level interval in which the remaining usable number of mold cycles falls is determined, and the corresponding level of mold remaining life assessment result is generated. At the same time, the remaining usable number of mold cycles, the corresponding life assessment level, and the current health status index are synchronously associated and stored to generate the corresponding mold remaining life assessment result.

[0042] After obtaining the mold's remaining life assessment results, the casting process parameters within the corresponding production cycle are further obtained. These parameters include at least the casting temperature, mold closing pressure, holding time, cooling time, and demolding interval. Subsequently, each casting process parameter is time-aligned according to a unified time axis, and corresponding process parameter change sequences are constructed. These process parameter change sequences are then correlated and matched with the health status index change sequence, and the parameter correlation coefficient between each casting process parameter and the degradation change slope is calculated. The parameter correlation coefficient is calculated by dividing the covariance between the process parameter change sequence and the degradation change slope change sequence by the corresponding standard deviation. The process involves: 1) After completing the traversal of all casting process parameters, extracting target process parameters whose correlation coefficients are higher than a preset correlation threshold, and statistically analyzing the changing trends of the corresponding target process parameters in different production cycles to determine the key process factors affecting mold life decay; 2) Synchronously integrating the mold comprehensive health status index curve, mold remaining life assessment results, mold remaining usable mold times data, key process factors, and corresponding process parameter changing trends, and generating a corresponding mold life assessment report according to a preset report template, wherein the mold life assessment report includes at least the current health status level, expected remaining service life, key degradation impact parameters, and maintenance early warning suggestions.

[0043] In summary, the embodiments of this application have at least the following technical effects: First, multi-dimensional state data of the casting mold is collected in real time and synchronously to obtain a multi-source heterogeneous state dataset. This dataset is then time-series aligned to generate an aligned dataset. Next, multi-scale degradation features are decoupled based on the aligned dataset to construct a multi-modal degradation feature vector. Then, an adaptive graph attention fusion network is constructed to fuse heterogeneous features of the multi-modal degradation feature vector, generating a comprehensive health status index curve for the mold. Finally, the remaining service life of the casting mold is predicted based on the comprehensive health status index curve. The remaining service life assessment results are correlated with casting process parameters to generate a mold service life assessment report. This solves the technical problem of accurately assessing the service life trend of casting molds during long-term casting operation in existing technologies, achieving the technical effect of improving the reliability of casting mold service life assessment.

[0044] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent assessment of casting mold life based on multi-source data analysis, characterized in that, The method includes: Real-time synchronous acquisition of multi-dimensional state data of casting molds, obtaining multi-source heterogeneous state datasets, performing time-series alignment, and generating aligned datasets; Based on the aligned dataset, multi-scale degradation features are decoupled to construct multi-modal degradation feature vectors; An adaptive graph attention fusion network is constructed to perform heterogeneous feature fusion on the multimodal degradation feature vectors to generate a comprehensive health status index curve for the mold. Based on the comprehensive health status index curve, the remaining service life of the casting mold is predicted, and the remaining service life assessment results of the mold are correlated with the casting process parameters to generate a mold service life assessment report.

2. The intelligent assessment method for casting mold life based on multi-source data analysis as described in claim 1, characterized in that, Real-time synchronous acquisition of multi-dimensional state data from casting molds, obtaining multi-source heterogeneous state datasets, performing time-series alignment, and generating an aligned dataset; methods include: Based on the multi-source sensor network configured on the casting production line, the surface marking image of the mold is acquired through the multi-source sensor network; Based on the identification image on the mold surface, the mold identification code is obtained through identity analysis. Synchronous monitoring is performed according to the mold identification code to obtain the mold closing trigger signal; The combined mode trigger signal is used as a global synchronization reference to synchronously collect and broadcast to the multi-source sensor network, driving the multi-source sensor network to trigger sampling and generate the multi-source heterogeneous state dataset.

3. The intelligent assessment method for casting mold life based on multi-source data analysis as described in claim 2, characterized in that, Methods for obtaining multi-source heterogeneous state datasets and performing time-series alignment to generate aligned datasets include: Based on the multi-source heterogeneous state dataset, command response delay analysis is performed to generate command response delay parameters; The time correction is performed on the multi-source heterogeneous state dataset according to the instruction response delay parameter to determine multiple data timestamps; The multi-source heterogeneous state dataset is divided according to the process stage window to generate multiple process segment data; The multiple process segment data are linearly time-aligned according to the multiple data timestamps to generate the aligned dataset.

4. The intelligent assessment method for casting mold life based on multi-source data analysis as described in claim 1, characterized in that, Based on the aligned dataset, multi-scale degradation features are decoupled to construct multi-modal degradation feature vectors. The method includes: The aligned dataset is traversed to perform parameter adaptive variational mode decomposition, generating a set of multi-scale intrinsic mode function components. The degradation sensitivity assessment is performed by traversing the set of multi-scale intrinsic mode function components, and the degradation sensitivity score is calculated. Based on the degradation sensitivity score, the alignment dataset is traversed to perform mold degradation matching and extract multiple sensitive components; The multiple sensitive components are divided into two-stage scales according to instantaneous frequency to generate a set of high-frequency degradation components and a set of low-frequency degradation components. Based on the set of high-frequency degradation components, cyclic feature capture is performed to extract local degradation features; Trend features are captured based on the set of low-frequency degradation components, and global degradation features are extracted. The local degradation features and the global degradation features are concatenated across modalities to generate the multimodal degradation feature vector.

5. The intelligent assessment method for casting mold life based on multi-source data analysis as described in claim 4, characterized in that, The method involves dividing the multiple sensitive components into a two-stage scale based on instantaneous frequency to generate a high-frequency degradation component set and a low-frequency degradation component set. Based on the multiple sensitive components, a Hilbert transform is performed to construct an instantaneous frequency sequence; Traverse the instantaneous frequency sequence to identify the median and determine the instantaneous frequency median; A mold opening and closing cycle is introduced, which includes a reference frequency, and sensitive components are extracted based on the reference frequency. The median instantaneous frequency is compared according to the mold opening and closing cycle: S1: The median of instantaneous frequencies higher than the sensitive components is included in the set of high-frequency degraded components; S2: Incorporate the median instantaneous frequency of the sensitive components that is lower than or equal to the median of the low-frequency degradation components into the low-frequency degradation component set.

6. The intelligent assessment method for casting mold life based on multi-source data analysis as described in claim 4, characterized in that, Based on the set of low-frequency degradation components, trend features are captured, and global degradation features are extracted. The method includes: A degradation trend curve is constructed by performing local weighted regression based on the set of low-frequency degradation components. The average monotonic slope of the degradation trend curve is obtained by calculation based on the degradation trend curve. A decline analysis was performed by mapping the average monotonic slope to the degradation trend curve, and the cumulative decline was extracted. The degradation trend curve is combined with the cumulative decrease rate to perform attenuation fitting, and the fitting structure parameters are extracted. The fitted structural parameters are added to the global degenerate feature.

7. The intelligent assessment method for casting mold life based on multi-source data analysis as described in claim 6, characterized in that, An adaptive graph attention fusion network is constructed to perform heterogeneous feature fusion on the multimodal degradation feature vectors to generate a comprehensive health status index curve for the mold. The method includes: The local degradation features and the global degradation features are each treated as independent feature nodes, forming a feature node set; The feature node set is traversed to calculate node mutual information and node correlation coefficient. Based on the node mutual information and node correlation coefficient, a weighted combination is performed to construct the connection strength value between nodes. The connection strength values ​​between the nodes are mapped to the feature node set for association matching to construct an initial fully connected feature relationship graph. The initial fully connected feature relationship graph is combined with the multimodal degradation feature vector and synchronized to the adaptive graph attention fusion network for heterogeneous feature fusion to generate the mold comprehensive health status index curve.

8. The intelligent assessment method for casting mold life based on multi-source data analysis as described in claim 7, characterized in that, The method involves synchronizing the initial fully connected feature graph with the multimodal degradation feature vector and then integrating it into the adaptive graph attention fusion network for heterogeneous feature fusion to generate the comprehensive health status index curve of the mold. An adaptive graph attention fusion network is constructed, which includes a multi-head graph attention layer, a temporal convolutional layer, and a feature aggregation layer. The multimodal degradation features are located according to the initial fully connected features to determine the real-time degradation stage features. The weights of multiple attention heads are dynamically adjusted according to the real-time degradation node features through the multi-head graph attention layer to generate stage adaptive fusion features. The temporal convolutional layer performs temporal smoothing on the stage adaptive fusion features to generate multiple node temporal features. The feature aggregation layer compresses and aggregates the multiple node temporal features to construct a one-dimensional degenerate feature sequence. The one-dimensional degradation feature sequence is subjected to degradation monotonicity constraints. Based on the constraint results, regression mapping is performed, and the monotonically decreasing curve is extracted as the comprehensive health status index curve of the mold.

9. The intelligent assessment method for casting mold life based on multi-source data analysis as described in claim 8, characterized in that, Based on the constraint results, regression mapping is performed to extract the monotonically decreasing curve. Methods include: Based on the constraint results, an order-preserving regression is performed on the zero-to-one interval to construct a nonparametric mapping function. The monotonically decreasing curve is extracted by reasoning based on the nonparametric mapping function.

10. The intelligent assessment method for casting mold life based on multi-source data analysis as described in claim 1, characterized in that, The remaining service life of the casting mold is predicted based on the comprehensive health status index curve to obtain the mold's remaining service life assessment result. The method includes: The comprehensive health status index curve is subjected to hybrid degradation trajectory prediction to generate a degradation trajectory prediction sequence; Based on the degradation trajectory prediction sequence, temporal positioning is performed to determine the failure prediction time value; The time difference between the real-time time value and the failure prediction time value is calculated to determine the remaining usable mold times. Based on the remaining available mold cycles data, synchronous quantification is performed to generate a mold remaining life assessment result.

Images