A lubrication alarm method for mold motion mechanism based on intelligent monitoring

Abstract

This invention belongs to the field of intelligent monitoring technology in industrial production, and relates to a lubrication alarm method for a mold motion mechanism based on intelligent monitoring. The method includes: acquiring sensor data of the mold motion mechanism in its operating state, and extracting features for each module to obtain a feature vector for each module; obtaining the gradient fluctuation degree of each module based on the continuous change of its feature vector; obtaining the local working condition clustering force index of each module based on the gradient fluctuation degree of each module; calculating the dynamic neighborhood sensing span of each module based on the local working condition clustering force index of each module; and improving the local anomaly factor algorithm based on the dynamic neighborhood sensing span of each module, thereby improving the accuracy of lubrication alarms for the mold motion mechanism.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring technology in industrial production, and relates to a lubrication alarm method for mold motion mechanism based on intelligent monitoring. It is used to monitor the lubrication status of the motion mechanism in the mold in real time and provide abnormal alarms, which can improve the accuracy of lubrication failure alarms. Background Technology

[0002] Currently, in modern injection molding and die casting industries, mold motion mechanisms such as sliders, angled ejectors, and ejector pins frequently undergo relative reciprocating motion. Without effective lubrication, this can easily lead to severe wear or even sintering and jamming of moving parts. To achieve non-invasive intelligent condition monitoring, time-series data anomaly detection technology based on acoustic or vibration characteristics is widely used. Vibration signals during mold operation can directly reflect changes in the friction state and lubrication conditions of moving pairs, providing a reliable monitoring basis for anomaly identification. Among these, the local anomaly factor algorithm, as a classic unsupervised learning technique, can identify anomalies based on local density deviations of data points, naturally possessing the potential to handle baseline drift under different operating conditions. This algorithm can complete model construction without relying on a large number of labeled samples, adapting to the monitoring needs of different mold models and production scenarios.

[0003] Chinese patent application CN111037987B discloses a signal processing method for monitoring the status of multi-link mechanisms in large mechanical presses. The method determines the mold replacement time by extracting signals in whole cycles, calculating similarity through dynamic time warping, and extracting the inherent frequency band of the oil film signal by wavelet packet decomposition to determine the oil supply status. This method enables the monitoring and diagnosis of the operating status and lubrication status of multi-link mechanisms. However, this scheme relies on the specific physical characteristics of the stamping process. When the working conditions of the mold motion mechanism change drastically, the distribution of vibration characteristics changes significantly, and the judgment benchmark based on fixed physical characteristics is prone to failure. Chinese patent application CN121434750B discloses a method for monitoring the operating status of mold control equipment. By denoising, normalizing and removing outliers from the mold operation dataset, a set of acoustic-vibration coupling modes is generated. The acoustic-vibration state feature set is obtained through feature index calculation and mode deviation assessment, and then the health status is calculated and the status is assessed. This scheme uses mode deviation assessment as the core means of anomaly detection. Essentially, it compares the current features with historical benchmarks and has a good monitoring effect on long-term steady-state working conditions. However, this scheme lacks the ability to adaptively process the differences in data distribution under different working conditions. Under sparse transitional working conditions, the deviation between the feature data and historical benchmarks naturally increases, which can easily misjudge normal working condition fluctuations as abnormal states and trigger frequent false alarms from the system.

[0004] In summary, existing local anomaly factor algorithms face serious robustness challenges when applied to mold lubrication monitoring: these algorithms heavily rely on pre-set fixed neighborhood parameters to assess local density. In actual mold production, the duration of different operating conditions varies significantly. Long steady-state production periods generate extremely dense feature data clusters, while short mold preheating or cycle adjustment transition periods generate extremely sparse small-scale data clusters. Fixed neighborhood parameters cannot accommodate the distribution of these different data cluster sizes. When the algorithm still forcibly searches according to a fixed number of maximum neighbors under sparse transition conditions, it leads to severe distortion in local density assessment, misjudging normal operating condition fluctuations as lubrication failures, thus triggering frequent false alarms from the system. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and solve the technical problem that the local anomaly factor algorithm in the prior art is prone to misjudging normal fluctuations as lubrication failures and causing frequent false alarms during the working condition transition period due to the use of fixed neighborhood parameters. The invention provides a lubrication alarm method for mold motion mechanism based on intelligent monitoring.

[0006] To achieve the above-mentioned objectives, this invention provides a lubrication alarm method for a mold motion mechanism based on intelligent monitoring, comprising the following steps: acquiring sensor data of the mold motion mechanism in operation and extracting module features to obtain feature vectors for each module; obtaining the gradient fluctuation degree of each module based on the continuous change of the feature vectors of each module; obtaining the local working condition clustering force index of each module based on the gradient fluctuation degree of each module; calculating the dynamic neighborhood sensing span of each module based on the local working condition clustering force index of each module; and improving the local anomaly factor algorithm based on the dynamic neighborhood sensing span of each module to realize alarm judgment for lubrication of the mold motion mechanism.

[0007] This method quantifies the severity of changes in working conditions by constructing a gradient fluctuation degree, and generates a local working condition clustering index based on this to characterize the clustering of feature vectors in multidimensional space. Then, based on this index, the neighborhood perception span of the local anomaly factor algorithm is dynamically adjusted so that the density assessment range adaptively matches the current data distribution pattern. This effectively avoids density calculation distortion caused by working condition fluctuations, accurately distinguishes between normal working condition fluctuations and real lubrication anomalies, significantly reduces the false alarm rate of the system, and improves the robustness and intelligence level of lubrication status monitoring of mold motion mechanisms.

[0008] The method for obtaining the feature vector for each module as described in this invention includes: deploying piezoelectric vibration sensors and Hall switches in the support column area of ​​the mold; setting the sampling frequency of the piezoelectric vibration sensors to 20kHz; using the level transition signal output by the Hall switch to mark the start and end times of a single mold opening and closing cycle; extracting the corresponding effective vibration data segments; after filtering the effective vibration data segments, calculating their root mean square value and peak factor; constructing the feature vector for the current module after normalization; and inputting the feature vector to the edge computing gateway in real time according to the module time sequence.

[0009] The present invention describes obtaining the gradient fluctuation level for each module, including: In the formula, Indicates the first The degree of gradient fluctuation in each module; Indicates the first Feature vectors of modulo order; Indicates the first Feature vectors of modulo order; Indicates the first The eigenvector of the i-th modulus and the i-th modulus Euclidean distance between eigenvectors of modulus n; For the first The 2-norm of eigenvectors of order n; Index of the modulus; Belongs to the natural constant An exponential function with base 1; To prevent constants with a denominator of 0.

[0010] This method normalizes the relative offset by dividing the Euclidean distance between adjacent modulus feature vectors by the L2 norm of the previous modulus feature vector, thus eliminating the impact of differences in feature vector amplitudes across different mold models or production batches. Simultaneously, an exponential function is introduced to amplify the normalized relative offset, enabling sensitive responses to changes in operating conditions. When the mold is in transitional periods such as preheating or cycle time adjustment, the gradient fluctuation increases exponentially, while when the mold enters a steady-state production period, the gradient fluctuation approaches zero. This provides accurate quantitative basis for calculating the local operating condition clustering index and adjusting the dynamic neighborhood sensing span in subsequent steps, ensuring that anomaly detection can adapt to the data distribution characteristics under different operating conditions.

[0011] The present invention describes obtaining the local working condition cohesion index for each module, including: In the formula, Indicates the first Local working condition cohesion index of each module; For the first The degree of gradient fluctuation in each module; Indicates the first Feature vectors of modulo order; This represents the feature vector of historical normal modules without alarms, extracted by the edge computing gateway through a sliding time window mechanism. This represents the total number of historical no-alarm, normal mode feature vectors. It is a natural constant; For the natural constant The logarithmic function with base 0; Indicates the first The eigenvector of the th modulus and the th eigenvector Euclidean distance between the eigenvectors of a historical normal mode without alarms; It is the index of the feature vector of the historical normal mode without alarms; Index of the modulus; Belongs to the natural constant An exponential function with base 1.

[0012] This method constructs a local operating condition clustering index to co-quantify the temporal fluctuation characteristics of the current module with its distribution in a multi-dimensional feature space. This index consists of two parts: the first part is a weighting factor based on the degree of gradient fluctuation. When the gradient fluctuation is large, i.e., during the transition period of the operating condition, the weighting factor is reduced to suppress the interference of unstable operating conditions on the spatial clustering assessment; the second part is an exponential decay term based on the historical normal mode eigenvector. This term measures the proximity of the current module's feature vector to the historical normal data distribution area; the closer the distance, the larger the term. Multiplying the two yields the local operating condition clustering index. A larger value indicates that the current module is in a steady-state production period and the feature clustering is stronger, while a smaller value indicates that it is in a transitional period or that the features deviate from the normal distribution. This provides a precise spatial distribution quantification basis for the subsequent adaptive adjustment of the dynamic neighborhood sensing span.

[0013] The total number of historical no-alarm normal mode feature vectors described in this invention ranges from 50 to 100.

[0014] The present invention describes obtaining the dynamic neighborhood sensing span for each module, including: In the formula, For the first The dynamic neighborhood perception span of each module; For the first Local working condition cohesion index of each module; This is a constant representing the maximum number of reference neighborhoods calculated by the system based on the total memory capacity limit of the historical feature data buffer; (symbol) This indicates that the result of the operation is rounded up. Belongs to the natural constant An exponential function with base 1.

[0015] This method constructs a dynamic neighborhood-aware span calculation model to map the local working condition clustering index to an adaptive neighborhood search range. When the local working condition clustering index is high, i.e., the mold is in a steady-state production period and the feature data is densely distributed, the negative exponential term... Approaching zero, the dynamic neighborhood sensing span is close to the maximum baseline neighborhood number constant, ensuring sufficient nearest neighbor nodes participate in density calculation in dense data clusters and improving the robustness of anomaly detection. When the cohesion index of a local working condition cluster is low, i.e. the mold is in the transition period of working condition and the feature data distribution is sparse, the negative exponent term increases, and the dynamic neighborhood sensing span is adaptively compressed, forcing the algorithm to search for nearest neighbors only in a very small local range, effectively avoiding density evaluation distortion caused by searching across working condition regions. The rounding up operation ensures that the neighborhood span is an integer, and the basis offset ensures that the algorithm has the minimum number of neighbor nodes required to construct the density topology structure. This mechanism enables the local anomaly factor algorithm to dynamically adjust the sensing range according to the actual data distribution state, fundamentally eliminating the problem of false alarms caused by fixed neighborhood parameters.

[0016] The maximum reference neighborhood number constant described in this invention has a value range of 20 to 30.

[0017] The improvement to the local anomaly factor algorithm described in this invention includes: directly replacing the default fixed neighbor search parameters of the local anomaly factor algorithm with dynamic neighborhood-aware span; the improved local anomaly factor calculation includes four steps in sequence: calculating the boundary distance, reachability distance, local reachability density, and local anomaly factor.

[0018] The boundary distance described in this invention is the spatial distance between the current module's feature vector and its farthest historical nearest neighbor node within its dynamic neighborhood. The reachability distance is the maximum value between the boundary distance of the current module's feature vector and the true Euclidean distance between the current module's feature vector and its historical nearest neighbor node for any historical nearest neighbor node within the dynamic neighborhood. The local reachability density is the arithmetic mean of the reachability distances of the current module's feature vector to all historical nearest neighbor nodes within its dynamic neighborhood, and the reciprocal of this arithmetic mean. The local anomaly factor is the ratio of the arithmetic mean of the local reachability densities of all historical nearest neighbor nodes within the dynamic neighborhood of the current module's feature vector to the local reachability density of the current module's feature vector itself.

[0019] This method improves the local anomaly factor algorithm by introducing a dynamic neighborhood sensing span: First, the boundary distance of the current module feature vector within its dynamic neighborhood is calculated; then, the maximum value between the historical nearest neighbor node boundary distance and the true Euclidean distance is used as the reachable distance to smooth numerical fluctuations; next, the arithmetic mean of the reachable distances of all nearest neighbor nodes in the dynamic neighborhood is calculated, and its reciprocal is taken as the local reachability density; finally, the ratio of the average local reachability density of all nearest neighbor nodes in the neighborhood to the local reachability density of the current node itself is used as the local anomaly factor. The larger this ratio is, the more abnormal the current module is, thereby achieving accurate identification of the lubrication state.

[0020] The present invention describes the alarm judgment for lubrication of mold motion mechanism, which includes: comparing the calculated local anomaly factor with the preset alarm judgment boundary in real time, wherein the value range of the alarm judgment boundary is 1.5 to 2.0; when the local anomaly factor is greater than the alarm judgment boundary, the edge computing gateway sends a lubrication interference alarm command to the production execution system and drives the on-site audio-visual prompt device to issue an alarm signal.

[0021] Compared with existing technologies, this invention offers at least the following advantages: First, by deploying piezoelectric vibration sensors and Hall switches, and combining them with mold opening and closing trigger signals, effective vibration data segments are precisely captured. The root mean square value and peak factor are extracted to construct feature vectors, thus reducing operational fluctuations and environmental noise interference from the data source, providing a stable and reliable data foundation for subsequent anomaly detection. Furthermore, a gradient fluctuation calculation model is constructed, normalizing the Euclidean distance between adjacent modulus feature vectors and amplifying it using an exponential function. This model sensitively responds to changes in operational conditions. During transitional periods such as mold preheating or cycle adjustment, the gradient fluctuation increases exponentially, approaching zero when entering steady-state production, providing a precise quantitative basis for distinguishing normal operational fluctuations from potential lubrication anomalies. Second, based on this, a local operational condition clustering index is constructed, combining the gradient fluctuation degree with the spatial distribution characteristics of historical normal feature vectors. During operational transition periods, the weight is automatically reduced to suppress unstable operational interference, and during steady-state periods, the spatial clustering of feature data is accurately reflected, providing a basis for subsequent adaptive adjustment of neighborhood parameters. The system provides precise spatial distribution quantification. Based on this index, a dynamic neighborhood-aware span calculation model is constructed, enabling the neighborhood search range of the local anomaly factor algorithm to automatically adjust with the data distribution status. In dense data clusters, the search range is expanded to ensure the robustness of density assessment, while in sparse data clusters, the search range is compressed to block cross-working-condition region searches, fundamentally solving the density assessment distortion problem caused by traditional fixed neighborhood parameters. Thirdly, the local anomaly factor algorithm is improved through dynamic neighborhood-aware span, ensuring that the calculation of boundary distance, reachable distance, local reachable density, and local anomaly factor matches the current working conditions. This effectively distinguishes between normal working-condition fluctuations and genuine lubrication anomalies, significantly reducing the system's false alarm rate and improving the accuracy of lubrication status monitoring of the mold motion mechanism. Finally, the calculated local anomaly factor is compared with the preset alarm judgment boundary in real time. When the threshold is exceeded, a lubrication interference alarm command is automatically triggered, realizing intelligent monitoring of the entire process from data acquisition, feature analysis, anomaly identification to lubrication alarm, providing reliable assurance for the continuous and stable operation of the mold motion mechanism. Attached Figure Description

[0022] Figure 1 This invention relates to a schematic flowchart of a lubrication alarm method for a mold motion mechanism based on intelligent monitoring.

[0023] Figure 2 This is a schematic diagram of the modulus eigenvalues ​​involved in the present invention.

[0024] Figure 3 This is a schematic diagram of the local working condition cohesion index involved in the present invention.

[0025] Figure 4 This is a schematic diagram of the dynamic neighborhood sensing span involved in the present invention.

[0026] Figure 5This is a schematic diagram of the improved LOF score alarm involved in the present invention.

[0027] Figure 6 This is a diagram showing the effect of this method compared to traditional methods. Detailed Implementation

[0028] The technical solution of the present invention will now be clearly and completely described in conjunction with the embodiments and accompanying drawings.

[0029] Example 1: This embodiment discloses a lubrication alarm method for a mold motion mechanism based on intelligent monitoring, referring to... Figure 1 This includes steps S1-S5: S1. Collect sensor data of the mold motion mechanism and extract mold features.

[0030] It should be noted that this step, through the coordinated design of sensor deployment and signal preprocessing, achieves stable acquisition and accurate extraction of friction state characteristics of the mold motion mechanism. This reduces feature distortion caused by operating condition fluctuations and environmental noise from the data source, ensuring that the feature data input to the anomaly detection model can truly reflect the lubrication state of the mold motion mechanism, and providing reliable data support for subsequent adaptive local anomaly factor calculation and lubrication state discrimination.

[0031] Specifically, piezoelectric vibration sensors and Hall switches are installed in the support column area of ​​the mold. The piezoelectric vibration sensors are used to collect vibration signals during the operation of the mold motion mechanism, and the Hall switches are used to collect trigger signals for the mold opening and closing actions. The sampling frequency of the piezoelectric vibration sensors is set to 20kHz to fully capture the weak high-frequency friction impact signals caused by lubrication abnormalities, while avoiding the introduction of excessive redundant data and reducing the edge computing load. The level transition signal output by the Hall switch is used to accurately mark the start and end times of a single mold opening and closing cycle. Based on the start and end times of the single mold opening and closing cycle, effective vibration data segments corresponding to the single mold opening and closing cycle are extracted from the continuous vibration acceleration signal to ensure that the extracted vibration data corresponds one-to-one with the actual operation process of the mold motion mechanism.

[0032] Furthermore, the effective vibration data segment corresponding to a single mold opening and closing cycle is filtered to remove random noise and environmental interference signals. The root mean square (RMS) value and peak factor of the vibration data segment are calculated respectively. The RMS value is used to characterize the overall energy distribution of the vibration signal, and the peak factor is used to reflect the severity of transient impacts in the vibration signal. After normalizing the RMS value and peak factor, the feature vector corresponding to the current mold cycle is constructed. The feature vector is then input to the edge computing gateway in real time according to the mold cycle time sequence to complete the standardized extraction and transmission of the single-cycle operation features of the mold motion mechanism.

[0033] Please see Figure 2 As shown, Figure 2 This is a schematic diagram of the modulus eigenvalues. As shown in the diagram, the modulus eigenvector contains two core indices, used to characterize different operating states of the mold's motion mechanism. The root mean square (RMS) normalized index primarily reflects the overall energy level of the vibration signal and can be used to quantify the severity of friction. The peak factor normalized index primarily reflects the intensity of transient impacts in the vibration signal and can be used to detect local friction anomalies caused by poor lubrication. Around modulus 100, 300, 650, and 900, both indices show significant peaks, corresponding to stages of mold preheating, process parameter adjustments, or deterioration of lubrication conditions, which highly matches actual production scenarios. During steady-state production, both indices remain at low and stable levels, indicating stable friction and good lubrication conditions in the mold's motion mechanism. After modulus 900, the RMS normalized index continues to climb and approaches its upper limit, while the peak factor normalized index fluctuates dramatically, accurately reflecting the physical process of deteriorating lubrication and increased friction in the mold's motion mechanism, providing reliable characteristic evidence for subsequent anomaly detection.

[0034] S2. Obtain the gradient fluctuation level for each module.

[0035] It should be noted that this step is used to quantify the degree of dynamic change of the time-series feature vector. By constructing the evolution gradient of the feature vectors of adjacent modules, the stability of the mold operation condition can be accurately identified, the steady-state production stage and the transition stage of the operation condition can be distinguished, and a basis can be provided for the dynamic adjustment of the adaptive neighborhood parameters in the future. This avoids abnormal misjudgments caused by the fluctuation of the operation condition and solves the problem that fixed parameters are difficult to adapt to the distribution of multi-condition data from the perspective of algorithm logic.

[0036] Specifically, considering the actual production and operation scenarios of molds, when a mold transitions from a cold preheating state to a steady-state production state, the frictional damping of moving parts such as sliders, inclined ejectors, and ejector pins in the mold's motion mechanism is prone to dynamic changes. During this process, the viscosity-temperature characteristics of the lubricating grease will change accordingly with temperature and motion load changes, which in turn will cause dynamic fluctuations in the motion impedance of the mold's motion mechanism. This change in the physical characteristics of the mold's motion mechanism will be directly mapped onto the extracted feature vectors of the time sequence modules. Specifically, the feature vectors of two consecutive adjacent modules will show a relative offset, and the magnitude of the offset is positively correlated with the severity of the change in operating conditions.

[0037] Furthermore, to accurately quantify the degree of drastic change in the current working condition of the mold and to clarify whether the feature vector offset is caused by fluctuations in normal working conditions, this step takes the feature vectors of two adjacent modules as the core calculation object. By combining normalized Euclidean distance with exponential mapping, a calculation model for the gradient fluctuation degree is constructed, and the gradient fluctuation degree of each module is obtained. The relationship is as follows: ; In the formula, Indicates the first The degree of gradient variability at the nth modulus is used to characterize the 1st modulus. The first module and the first The degree of drastic change in operating conditions between modules: the greater the gradient fluctuation, the more drastic the change in operating conditions between two adjacent modules; the smaller the gradient fluctuation, the more stable the operating conditions between two adjacent modules. Indicates the first Feature vectors of modulo order; Indicates the first Feature vectors of modulo order; Indicates the first The eigenvector of the i-th modulus and the i-th modulus The Euclidean distance between the feature vectors of each module is used to quantify the absolute offset of the feature vectors of two adjacent modules, and intuitively reflects the change in the friction state of the mold motion mechanism between adjacent modules. For the first The 2-norm of eigenvectors of order n; Index of the modulus; Based on the natural constant An exponential function with base 1; To prevent constants with a denominator of 0, for example, .

[0038] It should be noted that the normalization process is performed by dividing the Euclidean distance between the eigenvectors of adjacent moduli by the th moduli. The L2 norm of the feature vectors of each module converts the absolute offset of the feature vectors into relative offsets, effectively eliminating the influence of differences in feature vector amplitudes under different production batches and different mold models. This makes the gradient fluctuation level applicable to monitoring scenarios of mold motion mechanisms of different specifications and working conditions. The introduction of the exponential function can amplify the normalized relative offset, achieving a sensitive response to changes in working conditions.

[0039] Specifically, when the The eigenvector of the i-th modulus and the i-th modulus Euclidean distance between eigenvectors of each modulus An increase indicates that the mold is in a period of drastic transition, such as during machine preheating, process parameter adjustment, or production cycle switching. The gradient fluctuation is significant at this time. The gradient fluctuation increases exponentially, which clearly reflects the extreme instability of the physical state of the current mold motion mechanism. The shift in the eigenvector is due to normal operating condition fluctuations. When the relative displacement distance between the eigenvectors of adjacent mold cycles approaches zero, it indicates that the mold has entered a steady-state production period after thermal equilibrium, and the friction and lubrication states of the mold motion mechanism tend to stabilize. At this time, the gradient fluctuation degree... The synchronization value approaches zero, which can accurately represent that the working condition is in a highly stable dwell state and the fluctuation of the feature vector is within the normal range. At the same time, the gradient fluctuation degree of each module is calculated and transmitted synchronously with the feature vector of the corresponding module to the edge computing gateway. This provides a real-time and reliable quantitative basis for the adaptive adjustment of the neighborhood parameters of the local anomaly factor algorithm in subsequent steps, ensuring that subsequent anomaly detection can adapt to the feature data cluster distribution of different working conditions and further reduce the probability of false alarms.

[0040] S3. Obtain the local working condition cohesion index for each module.

[0041] It should be noted that the core of this step is to construct a local operating condition clustering index by combining the temporal fluctuation characteristics and the spatial distribution characteristics of historical data. This index is specifically used to accurately characterize the natural spatial compactness of the feature vector of the current module in the multi-dimensional feature space, making up for the deficiency that the gradient fluctuation degree alone cannot fully reflect the overall distribution of feature data. It provides an accurate and reliable spatial distribution quantification basis for the dynamic adaptation of neighborhood parameters in the subsequent local anomaly factor algorithm, ensuring that all terms in the entire method are completely consistent, while taking into account the real-time performance and accuracy of the calculation, adapting to the processing capabilities of the edge computing gateway, and ensuring the accuracy of subsequent anomaly detection.

[0042] Specifically, in the multidimensional feature space, assessing the degree of temporal change solely through the degree of gradient fluctuation cannot fully characterize the spatial distribution of the feature vectors of the current module. The feature data generated at different operating stages exhibit significant and distinguishable differences in their distribution patterns within the multidimensional feature space. The feature data generated during the transition phase, due to the unstable operating conditions of the mold, the continuous fluctuations in the friction state of moving parts, and the dynamic adjustment of the lubrication state, presents a loose, banded distribution in the multidimensional feature space, with strong dispersion of feature data points, making it impossible to form compact clusters. In contrast, the feature data generated during the steady-state production phase, due to the stable operating conditions of the mold, the stable friction state of moving parts, and the good lubrication state, presents a compact, clustered distribution in the multidimensional feature space, with strong aggregation of feature data points, forming obvious dense data clusters.

[0043] Furthermore, in order to accurately assess the natural spatial compactness attribute near the current data point by combining the current distribution of historical data, a local working condition clustering index is constructed for each module, with the following relationship: ; In the formula, Indicates the first The local working condition cohesion index of the first module is used to characterize the first... The degree of local clustering of the feature vectors of each module in the multidimensional feature space directly corresponds to the spatial compactness of the feature data points. The larger the value, the stronger the clustering of the feature vector of the current module with the surrounding feature data points, which corresponds to a more stable mold operating condition and a better lubrication state. The smaller the value, the stronger the dispersion of the feature vector of the current module with the surrounding feature data points, which corresponds to a more unstable mold operating condition and may be in a transitional stage. For the first The gradient fluctuation of each module is used to reflect the degree of change in working conditions between the current module and the previous module. Its value directly affects the weighting of the local working condition cohesion index, so as to achieve a coordinated characterization of the temporal change degree and the spatial distribution state. Indicates the first The feature vectors of each module are used to characterize the friction and lubrication states of the current module's motion mechanism. This represents the historical no-alarm normal module feature vector extracted by the edge computing gateway through a sliding time window mechanism. All historical no-alarm normal module feature vectors come from modules in normal operation without triggering any lubrication alarms, ensuring the validity and reference value of historical feature data and providing a reliable historical benchmark for evaluating the spatial distribution of current module feature vectors. This represents the total number of historical no-alarm, normal mode feature vectors. It is a natural constant; For the natural constant The logarithmic function with base 0; Indicates the first The eigenvector of the th modulus and the th eigenvector The Euclidean distance between the historical normal feature vectors without alarms is used to quantify the spatial deviation between the current feature vector and the historical normal feature vectors. The value is positively correlated with the degree of spatial deviation. The larger the value, the farther the current feature vector deviates from the historical normal feature data distribution area; the smaller the value, the closer the current feature vector is to the historical normal feature data distribution area. It is the index of the feature vector of the historical normal mode without alarms; Index of the modulus; Belongs to the natural constant An exponential function with base 1.

[0044] It should be noted that the natural constant is introduced. The purpose is to avoid the calculation failure caused by the logarithmic denominator term being zero when the gradient fluctuation is zero, to ensure the universality and stability of the local working condition clustering index calculation model, and to ensure that the calculation can be completed normally under any working condition.

[0045] Specifically, the total number of historical no-alarm normal mode feature vectors The preferred value range is 50 to 100. This range has been verified through multiple experiments, taking into account both the statistical representativeness of historical benchmarks and the reasonableness of computational costs; when the total number Below At that time, the number of historical no-alarm normal mode feature vectors extracted is insufficient, the statistical representativeness of the historical benchmark is weak, and it cannot accurately reflect the spatial distribution pattern of feature data under normal operating conditions. This can easily lead to distortion of the local operating condition clustering index calculation results, affecting the accuracy of subsequent anomaly detection; when the total number Higher than This not only significantly increases the computational overhead of the edge computing gateway and reduces the real-time performance of feature extraction, index calculation, and subsequent alarm judgment, but also easily introduces early-failed normal module feature vectors. The operating status of the module corresponding to these early feature vectors may have undergone subtle changes, making it impossible to accurately serve as a benchmark for evaluating the spatial distribution of the current module feature vectors, thereby interfering with the calculation accuracy of the local working condition clustering index.

[0046] Furthermore, when the first Gradient variability of each mode When the value increases, it indicates that the current mold is in a non-steady-state transition phase, the mold's operating conditions are unstable, and the friction state of moving parts fluctuates drastically. At this time, the weight term... The value of the denominator will decrease as the logarithmic term increases. The physical meaning of this is that during the transition phase of the operating condition, the spatial clustering of the feature data itself has no reference value. By reducing the value of the preceding weight term, the interference of the temporal unstable state on the local operating condition clustering index is reduced, and the misjudgment of the index caused by the transition of the operating condition is avoided.

[0047] Specifically, because the feature data during the transition phase inevitably deviates from the dense distribution area of ​​historical alarm-free normal modal feature data, the Euclidean distance between the current modal feature vector and each historical alarm-free normal modal feature vector will increase significantly, leading to a sharp decrease in the subsequent exponential decay term in the calculation formula; the combined effect of these two factors makes the first... Local working condition clustering index of each module The gradient fluctuation decreases significantly; conversely, when the gradient fluctuation decreases, the weight term increases and the magnitude of the decrease in the exponential decay term decreases, and the clustering index remains at a high level. The combined effect of the two results in a significant increase in the local working condition clustering index of the module.

[0048] Please see Figure 3As shown, Figure 3 This is a schematic diagram of the local operating condition clustering index. As shown in the diagram, the local operating condition clustering index can accurately quantify the degree of local clustering of the feature vectors of each module in the multidimensional feature space. Its value shows a significant positive correlation with the stability of the mold's operating conditions. In the initial module stage, the local operating condition clustering index remains at its highest level, corresponding to the brief stable phase before the mold enters a stable operating state at the beginning of startup. As the module progresses, the local operating condition clustering index shows a significant sharp drop in the intervals corresponding to the peak values ​​of the feature vectors, forming significant troughs. This accurately reflects the loose distribution and strong dispersion of feature data in the multidimensional space during the transition phase of the operating conditions, which is consistent with non-steady-state processes such as cold machine preheating and process parameter adjustments. The physical characteristics of the conditions are highly consistent. In most of the mold production cycles, the local working condition clustering index remains at a high level with small fluctuations, indicating that the feature data in this stage presents a compact cluster distribution in multidimensional space, and the friction and lubrication states of the mold motion mechanism are stable. However, in the interval after the 800th cycle, the local working condition clustering index shows a continuous and significant decrease, and the frequency of fluctuations increases significantly. This corresponds to a stage where the feature vector changes drastically, fully demonstrating the evolution law that the feature data gradually deviates from the historical normal dense area and the spatial clustering continues to decrease during the deterioration of the lubrication state. This provides a precise and effective quantitative basis for the adaptive adjustment of the subsequent dynamic neighborhood perception span.

[0049] S4. Obtain the dynamic neighborhood perception span for each module.

[0050] It should be noted that this step achieves adaptive adjustment of the neighborhood range through the local working condition clustering index, generating a dynamic neighborhood perception span that matches the current data distribution pattern; by dynamically constraining the search range of density calculation, it avoids abnormal cross-working condition search and density evaluation collapse in sparse data scenarios, thereby improving the stability and accuracy of mold motion mechanism lubrication status identification under variable working conditions.

[0051] Specifically, to ensure that the density assessment range is consistent with the data aggregation level of the current module, a reasonable search range suitable for the current feature points is obtained based on the local working condition clustering index mapping. A dynamic neighborhood-aware span calculation model is then constructed with this as the objective, and the 1st... The dynamic neighborhood sensing span corresponding to each module is expressed by the following formula: ; In the formula, For the first The dynamic neighborhood perception span of each module; For the first Local working condition cohesion index of each module; This is a constant representing the maximum number of reference neighborhoods calculated by the system based on the total memory capacity limit of the historical feature data buffer; (symbol) This represents performing a floor operation on the result of the operation; numbers As a base offset, it is used to ensure that the algorithm has the minimum number of neighbor nodes required to build a density topology; Belongs to the natural constant An exponential function with base 1.

[0052] In this embodiment, the maximum reference neighborhood number constant The preferred value range is 20 to 30. This range can balance the robustness of density assessment under steady-state conditions with the actual computing power limitations of edge computing gateways, ensuring detection accuracy while maintaining the real-time operation capability of the system.

[0053] When the When the local cohesion index of a single mold is high, meaning the mold is in a long-term thermally stable production period, the negative index term... It rapidly approaches zero; at this point, the first... The dynamic neighborhood sensing span of the first modulus adaptively expands and approaches the maximum baseline neighborhood size constant, ensuring sufficient contrast nodes in dense data clusters to maintain the robustness of density assessment; conversely, when the first modulus... Local working condition clustering index of each module When the temperature drops sharply, the operating condition enters a period of fluctuation. The dynamic neighborhood perception span of each module is forcibly compressed proportionally, forcing subsequent algorithms to focus only on feature nodes within their own very small range. This mechanism can effectively block invalid searches across working conditions, fundamentally avoid density calculation distortion caused by excessively large search ranges, and eliminate false anomaly judgments caused by working condition fluctuations.

[0054] Please see Figure 4 As shown, Figure 4This is a schematic diagram of the dynamic neighborhood sensing span. As shown in the diagram, the dynamic neighborhood sensing span can adaptively adjust the neighborhood range in real time based on the changes in the local working condition clustering index. Its numerical changes show a high degree of matching with the distribution pattern of the mold's operating conditions. The dynamic neighborhood sensing span is always constrained within the range of the maximum baseline neighborhood number, and there is no expansion beyond this upper limit, which fully complies with the algorithm's restriction rules on the neighborhood search range. When the mold is in the steady-state production stage, the dynamic neighborhood sensing span remains at a high level and continuously approaches the maximum baseline neighborhood number. This enables the local anomaly factor algorithm to select sufficient comparison nodes in dense data clusters, ensuring the noise resistance and robustness of local density assessment. During the transition phase of the operating condition, in the interval corresponding to the sharp drop in the local operating condition cluster cohesion index, the dynamic neighborhood perception span undergoes significant step-like compression, and the neighborhood search range is strictly limited to a very small range. In the early stage of lubrication deterioration after module 800, the neighborhood span shrinks synchronously with the decrease in cluster cohesion index, ensuring that anomaly detection is carried out only around the local feature points of the current operating condition, effectively improving the accuracy of subsequent lubrication anomaly identification.

[0055] S5. Improve the local anomaly factor algorithm to realize mold lubrication anomaly alarm.

[0056] It should be noted that this step adaptively improves the traditional local anomaly factor algorithm through dynamic neighborhood perception span, ensuring that the local density calculation process matches the actual data distribution of the current module. The dynamically adjusted anomaly detection mechanism effectively distinguishes between operating condition fluctuations and genuine lubrication anomalies, improving the accuracy of lubrication status determination. Based on the calculated local anomaly factors, threshold comparison and alarm output are completed, forming a complete intelligent monitoring and lubrication alarm closed loop, ensuring the stability and reliability of the mold's motion mechanism during continuous operation.

[0057] Specifically, for the feature vector corresponding to the current module, the dynamic neighborhood-aware span is used to directly replace the default fixed neighbor search parameters of the local anomaly factor algorithm, enabling the density assessment and anomaly identification process to have adaptive capabilities. The improvement of the local anomaly factor algorithm includes four steps: calculating the boundary distance, reachability distance, local reachability density, and local anomaly factor. First, taking the feature vector of the current module as the target object, the nearest neighbor search range is determined according to the dynamic neighborhood-aware span. The nearest neighbor nodes of the current feature vector are selected from the feature vectors of historical normal modules without alarms and are denoted as the historical nearest neighbor node set. The spatial distance from the current feature vector to the farthest historical nearest neighbor node within its corresponding dynamic neighborhood range is calculated, and this distance is defined as the boundary distance of the current feature vector. Second, for any historical nearest neighbor node in the historical nearest neighbor node set of the current feature vector, the current module... The reachability distance from the current modulus feature vector to its historical nearest neighbor is determined by the larger of the boundary distance of the current modulus feature vector and the true Euclidean distance between the current modulus feature vector and its historical nearest neighbor. Then, all historical nearest neighbors within the dynamic neighborhood corresponding to the current modulus feature vector are extracted. The arithmetic mean of the reachability distances from the current modulus feature vector to all these historical nearest neighbors is calculated. The reciprocal of this arithmetic mean is then used as the local reachability density of the current modulus feature vector itself. If this arithmetic mean is 0, meaning the reachability distances of all nearest neighbors are 0, the local reachability density of the current feature vector is set to 0. Finally, the local reachability densities corresponding to all nearest neighbors within the dynamic neighborhood are obtained and their average value is calculated. This average value is then compared with the local reachability density of the current feature vector itself to obtain the local anomaly factor corresponding to the current modulus.

[0058] Furthermore, the local anomaly factor is compared with the preset alarm judgment boundary in real time. The value range of the alarm judgment boundary is set from 1.5 to 2.0. When the local anomaly factor is greater than the alarm judgment boundary, it indicates that the operating status of the mold motion mechanism is significantly abnormal compared with the current local working condition neighborhood, and the lubrication status shows a significant deterioration trend. After the edge computing gateway identifies the abnormal status, it immediately sends a lubrication interference alarm command to the production execution system, and at the same time drives the on-site audible and visual prompt device to issue an alarm signal to remind the user to perform lubrication maintenance operations. This completes the complete intelligent monitoring closed loop of the mold motion mechanism from data acquisition, feature analysis, anomaly identification to lubrication alarm.

[0059] Please see Figure 5 As shown, Figure 5This is a schematic diagram of the improved LOF score alarm. As shown in the diagram, the improved local anomaly factor algorithm can accurately distinguish between operating condition fluctuations and actual lubrication anomalies, achieving stable and reliable alarm output. During most of the steady-state production stages of mold operation, the local anomaly factor score remains below the alarm threshold with small fluctuations, without triggering any alarms, effectively avoiding misjudging normal operating condition fluctuations as lubrication failures. During the operating condition transition stage, although the local anomaly factor score fluctuates to some extent, due to the adaptive adjustment of the dynamic neighborhood perception span, the algorithm only performs density evaluation within a local range, so the score does not continuously exceed the alarm threshold, and no false alarms are generated. However, in the lubrication deterioration stage after 900 cycles, the local anomaly factor score rises significantly and remains above the alarm threshold, triggering continuous alarm signals, accurately capturing the physical process of deterioration of friction state and lubrication conditions in the mold motion mechanism.

[0060] Please see Figure 6 As shown, Figure 6 This diagram illustrates the performance comparison between the proposed method and traditional methods. As shown, there are significant differences in the trends of local anomaly factor scores between the two methods. The proposed method demonstrates a clear technical advantage over the traditional fixed-neighborhood method in terms of adaptability to operating conditions and accuracy of anomaly identification. Within most of the mold's steady-state production intervals, the improved method exhibits smaller fluctuations in local anomaly factor scores, remaining within a relatively stable range. In contrast, the traditional method frequently shows significant jumps in scores, with the jumps far exceeding actual operating condition requirements. During transitional phases such as cold-machine preheating and process parameter adjustments, the traditional method repeatedly exceeds the alarm threshold value range. The proposed method, relying on the adaptive adjustment capability of the dynamic neighborhood sensing span, only shows small score fluctuations during the same transitional phases, without significant increases in anomaly scores across operating conditions. Within the mold lubrication deterioration intervals, the improved method shows a continuous and stable upward trend in local anomaly factor scores. It can accurately pinpoint density deviations caused by lubrication anomalies; while traditional methods exhibit irregular and violent oscillations in this range, failing to establish a stable anomaly identification benchmark and effectively distinguishing between lubrication degradation and normal operating condition fluctuations. Overall, the improved method achieves deep coupling between the calculation process of local anomaly factors and the actual operating conditions of the mold through precise control of the dynamic neighborhood perception span. This completely solves the anomaly identification distortion problem caused by fixed neighborhood parameters in traditional methods. While ensuring effective detection of lubrication anomalies, it significantly reduces the probability of false identification under normal operating conditions, fully verifying the practicality and superiority of the method in the intelligent monitoring and alarm application of lubrication status of mold motion mechanisms.

Claims

1. A lubrication alarm method for a mold motion mechanism based on intelligent monitoring, characterized in that, The main process steps include: Acquire sensor data of the mold motion mechanism in operation, and extract the feature vector of each module. Based on the continuous changes in the feature vector of each module, the gradient fluctuation degree of each module is obtained; Based on the gradient fluctuation degree of each module, the local working condition cohesion index of each module is obtained; Based on the local working condition clustering index of each module, the dynamic neighborhood sensing span of each module is calculated. Based on the dynamic neighborhood sensing span of each module, the local anomaly factor algorithm is improved, and alarm judgment for lubrication of the mold motion mechanism is realized.

2. The lubrication alarm method for a mold motion mechanism based on intelligent monitoring according to claim 1, characterized in that, The process of obtaining the feature vector for each module includes: Piezoelectric vibration sensors and Hall switches are deployed in the support column area of ​​the mold. The sampling frequency of the piezoelectric vibration sensors is set to 20kHz. The level transition signal output by the Hall switch is used to mark the start and end times of a single mold opening and closing cycle, and the corresponding effective vibration data segments are extracted. After filtering the effective vibration data segments, their root mean square value and peak factor are calculated. After normalization, the feature vector of the current module is constructed. The feature vector is then input to the edge computing gateway in real time according to the module time sequence.

3. The lubrication alarm method for a mold motion mechanism based on intelligent monitoring according to claim 1, characterized in that, The process of obtaining the gradient fluctuation level for each module includes: ； In the formula, Indicates the first The degree of gradient fluctuation in each module; Indicates the first Feature vectors of modulo order; Indicates the first Feature vectors of modulo order; Indicates the first The eigenvector of the i-th modulus and the i-th modulus Euclidean distance between eigenvectors of modulus n; For the first The 2-norm of eigenvectors of order n; Index of the modulus; Belongs to the natural constant An exponential function with base 1; To prevent constants with a denominator of 0.

4. The lubrication alarm method for a mold motion mechanism based on intelligent monitoring according to claim 1, characterized in that, The process of obtaining the local working condition cohesion index for each module includes: ； In the formula, Indicates the first Local working condition cohesion index of each module; For the first The degree of gradient fluctuation in each module; Indicates the first Feature vectors of modulo order; This represents the feature vector of historical normal modules without alarms, extracted by the edge computing gateway through a sliding time window mechanism. This represents the total number of historical no-alarm, normal mode feature vectors. It is a natural constant; For the natural constant The logarithmic function with base 0; Indicates the first The eigenvector of the th modulus and the th eigenvector Euclidean distance between the eigenvectors of a historical normal mode without alarms; It is the index of the feature vector of the historical normal mode without alarms; Index of the modulus; Belongs to the natural constant An exponential function with base 1.

5. A lubrication alarm method for a mold motion mechanism based on intelligent monitoring according to claim 4, characterized in that, The total number of historical no-alarm normal mode feature vectors ranges from 50 to 100.

6. The lubrication alarm method for a mold motion mechanism based on intelligent monitoring according to claim 1, characterized in that, The process of obtaining the dynamic neighborhood-aware span for each module includes: ； In the formula, For the first The dynamic neighborhood perception span of each module; For the first Local working condition cohesion index of each module; This is a constant representing the maximum number of reference neighborhoods calculated by the system based on the total memory capacity limit of the historical feature data buffer; (symbol) This indicates that the result of the operation is rounded up. Belongs to the natural constant An exponential function with base 1.

7. A lubrication alarm method for a mold motion mechanism based on intelligent monitoring according to claim 6, characterized in that, The value of the maximum reference neighborhood number constant ranges from 20 to 30.

8. A lubrication alarm method for a mold motion mechanism based on intelligent monitoring according to claim 1, characterized in that, The improvement to the local anomaly factor algorithm includes: directly replacing the default fixed neighbor search parameters of the local anomaly factor algorithm with dynamic neighborhood-aware span. The improved local anomaly factor calculation includes four steps: calculating the boundary distance, reachability distance, local reachability density, and local anomaly factor.

9. A lubrication alarm method for a mold motion mechanism based on intelligent monitoring according to claim 8, characterized in that, The boundary distance is the spatial distance between the current module's feature vector and its farthest historical nearest neighbor node within its dynamic neighborhood. The reachability distance is the maximum value between the boundary distance of the current module's feature vector and the true Euclidean distance between the current module's feature vector and its historical nearest neighbor node for any historical nearest neighbor node within the dynamic neighborhood. The local reachability density is the arithmetic mean of the reachability distances of the current module's feature vector to all its historical nearest neighbor nodes within its dynamic neighborhood, and the reciprocal of this arithmetic mean. The local anomaly factor is the ratio of the arithmetic mean of the local reachability densities of all historical nearest neighbor nodes within the dynamic neighborhood of the current module's feature vector to the local reachability density of the current module's feature vector itself.

10. A lubrication alarm method for a mold motion mechanism based on intelligent monitoring according to claim 1, characterized in that, The method for realizing alarm judgment of lubrication of mold motion mechanism includes: comparing the calculated local anomaly factor with the preset alarm judgment boundary in real time, wherein the value range of the alarm judgment boundary is 1.5 to 2.0; when the local anomaly factor is greater than the alarm judgment boundary, the edge computing gateway sends a lubrication interference alarm command to the production execution system and drives the on-site audio-visual prompt device to issue an alarm signal.

Images