A safety warning method for powder alloy processing equipment

Abstract

The application relates to the technical field of monitoring and early warning, in particular to a safety early warning method for a powder alloy processing device, which comprises the following steps: collecting device state data and vectorizing the feature vectors; for each data point in a health state reference data set, labeling an inherent stability weight based on the historical fluctuation degree thereof; for a to-be-tested data point acquired in real time, determining a dynamic neighborhood parameter based on the locally sparseness of the to-be-tested data point after stability weighting, combining the instantaneous fluctuation degree of the to-be-tested data point to apply a risk penalty, and calculating a local anomaly factor of the to-be-tested data point based on the dynamic neighborhood parameter to determine whether the device enters a sub-health state. According to the technical scheme, weak early anomalies can be recognized from the neighborhood structure level of data points, and the accuracy and sensitivity of the device sub-health state early warning are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of monitoring and early warning, and particularly relates to a safety early warning method for a powder alloy processing device. BACKGROUND

[0002] In the production process of high-performance powder alloys, the stability of various processing equipment, such as gas atomization powder production equipment, high-energy ball mills, and vibrating screening machines, is the core prerequisite for ensuring the quality of the final product and the safety of the production process. Before a serious failure occurs, the equipment usually goes through a sub-healthy stage of gradually deteriorating performance, during which the physical signals such as vibration and acoustics of the equipment will change weakly but can be identified.

[0003] Currently, in order to capture early-stage abnormalities of the equipment, unsupervised anomaly detection algorithms such as Local Outlier Factor (LOF) have been applied in the prior art. The basic idea of this kind of algorithm is to convert the running state of the equipment at different times into data points in the feature space through sensor data, and identify isolated points or outliers that deviate from the normal data cluster through data mining techniques, thereby realizing fault warning.

[0004] However, the unsupervised anomaly detection algorithm such as LOF mainly focuses on the absolute position of the data point in the feature space, but the sub-healthy state of the equipment, especially in the early stage of failure, may have a very close spatial position to the normal point, and does not show significant outlier characteristics. At this time, the anomaly is more reflected in the distortion of the micro data environment structure around the data point, rather than the spatial isolation of the data point itself. Although some existing improved methods, such as dynamically adjusting the neighborhood parameter k of the LOF algorithm, improve the adaptability of the algorithm to a certain extent, these improvements still stay within the original feature space to optimize the algorithm locally, and when the abnormal state does not show obvious spatial distance anomaly in the original feature space, the detection ability will be essentially limited or even failed. SUMMARY

[0005] In order to solve the technical problem of missing detection caused by simply relying on the spatial position of the data point for anomaly detection in the prior art, the present application provides a safety early warning method for a powder alloy processing device.

[0006] This application provides a safety early warning method for powder alloy processing equipment, comprising: collecting operating status data of the powder alloy processing equipment to generate a series of multi-dimensional feature vectors, and constructing a health status reference dataset based on the multi-dimensional feature vectors; labeling the reference data points with inherent stability weights that characterize their intrinsic stability based on the fluctuation of historical feature vectors corresponding to each reference data point in the health status reference dataset; acquiring test data points in real time, evaluating the local sparsity of the data environment in which the test data points are located based on the inherent stability weights of the neighboring data points of the test data points, and applying risk penalties in combination with the instantaneous fluctuation of the test data points themselves, so as to determine dynamic neighborhood parameters suitable for the test data points; calculating the local anomaly factor of the test data points based on the dynamic neighborhood parameters, and determining whether the powder alloy processing equipment has entered a sub-healthy state based on the local anomaly factor, thereby realizing a safety early warning for the powder alloy processing equipment.

[0007] This application combines stability-weighted local sparsity assessment with risk penalty for instantaneous fluctuations to dynamically determine the optimal analysis scale for each data point to be measured, thereby achieving a combination of environmental judgment and risk perception of equipment status. This not only effectively suppresses false alarms during equipment switching from normal operating conditions, but also enables a rapid and sensitive response to sudden early fault symptoms.

[0008] In one embodiment, the inherent stability weights satisfy the following relationship: ;in, Reference data points The inherent stability weights, The reference data point The historical volatility corresponding to the fluctuation of the historical eigenvector. and These represent the minimum and maximum values ​​of all historical volatility in the health status reference dataset, respectively. This is the preset weight decay coefficient.

[0009] It can map the historical volatility, which represents the amplitude of state fluctuations, into a weight that reflects the reliability of data points as healthy samples. The decreasing exponential function can map the weight of stable states to close to 1, while applying a non-linear, fast penalty to the weight of unstable states, causing it to quickly approach 0, thereby accurately labeling the credibility of each healthy data point.

[0010] In one embodiment, the evaluation process of local sparsity involves determining the initial neighborhood of the data point to be tested and calculating the stability-weighted average distance between the data point to be tested and each of its neighbors within the initial neighborhood. The stability-weighted average distance satisfies the following relationship: ;in, a stability weighted average distance of the to-be-tested data point a stability weighted average distance of the to-be-tested data point a size of an initial neighborhood, an initial neighborhood of the to-be-tested data point an initial neighborhood of the to-be-tested data point an initial neighborhood of the to-be-tested data point an initial neighborhood of the to-be-tested data point an initial neighborhood of the to-be-tested data point an initial neighborhood of the to-be-tested data point an initial neighborhood of the to-be-tested data point

[0011] By taking the inherent stability weight of the neighbor point as the divisor of the spatial distance, the dynamic adjustment of the distance is realized, so that the evaluation of the local sparsity no longer treats all neighbors equally, but rather focuses more on neighbor points from reliable working conditions, thereby effectively filtering out the interference caused by normal working condition fluctuations and obtaining a more accurate judgment of the data environment.

[0012] In one embodiment, the dynamic neighborhood parameter satisfies the relationship: ; wherein, the dynamic neighborhood parameter, a reference neighborhood parameter determined based on the evaluation result of the local sparsity, a lower limit of the neighborhood parameter, an instantaneous fluctuation degree corresponding to the instantaneous fluctuation of the to-be-tested data point itself, a fluctuation penalty coefficient, a floor function.

[0013] When the equipment runs smoothly, the instantaneous fluctuation degree is small, the penalty term is close to 0, and the dynamic neighborhood parameter is mainly determined by the data environment. When the equipment state appears a severe fluctuation, the instantaneous fluctuation degree is large, and the penalty term is close to 1, which will forcibly and significantly reduce the neighborhood parameter, so as to tighten the detection scale and switch to a high sensitivity mode, thereby ensuring a high sensitivity to sudden deterioration.

[0014] In one embodiment, the multi-dimensional feature vector includes time domain statistical features and frequency domain statistical features; the time domain statistical features include root mean square, kurtosis and skewness; and the frequency domain statistical features include dominant frequency and spectral entropy.

[0015] In one embodiment, the historical fluctuation degree is calculated by: obtaining a preset number of historical feature vectors before the reference data point, calculating a geometric centroid of the historical feature vectors, and calculating an average Euclidean distance from the historical feature vectors to the geometric centroid.

[0016] In one embodiment, the calculation of the instantaneous fluctuation degree is as follows: a preset number of historical feature vectors before the data point to be tested are obtained, a geometric centroid of the historical feature vectors is calculated, and an average Euclidean distance of the historical feature vectors to the geometric centroid is calculated.

[0017] In one embodiment, the reference neighborhood parameter satisfies the relationship: ; wherein, is a preset upper limit of the neighborhood parameter, is a sparsity index generated according to the local sparsity evaluation, is a sparsity response coefficient.

[0018] In one embodiment, the local anomaly factor of the data point to be tested is calculated based on the dynamic neighborhood parameter, specifically, the local anomaly factor score of the data point to be tested is calculated by using a local anomaly factor algorithm.

[0019] In one embodiment, the determination of whether the powder alloy processing equipment enters a sub-healthy state according to the local anomaly factor includes: comparing the calculated local anomaly factor of the data point to be tested with a preset warning threshold; if the local anomaly factor is greater than the warning threshold, it is determined that the powder alloy processing equipment enters a sub-healthy state and a warning signal is triggered; if the local anomaly factor is not greater than the warning threshold, it is determined that the powder alloy processing equipment is in a healthy state.

[0020] The technical solution of the present application has the following beneficial technical effects:

[0021] The present application can combine the stability-weighted local sparsity evaluation and the risk penalty of the instantaneous fluctuation degree to dynamically determine the optimal analysis scale for each data point to be tested, realize the combination of environmental judgment and risk perception of the equipment state, effectively suppress false alarms during normal working condition switching, and quickly and sensitively respond to sudden early signs of failure.

[0022] Further, the stability-weighted greatly enhances the model's ability to recognize normal working condition switching, effectively suppresses false alarms, and through the instantaneous fluctuation penalty, gives the system risk perception ability, realizes high sensitivity to sudden deterioration, not only can detect gradual drift, but also can quickly and sensitively respond to sudden early signs of failure, and comprehensively improves the accuracy and reliability of the warning. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 is a flowchart of a safety warning method for a powder alloy processing equipment according to an embodiment of the present application.

[0024] Figure 2is a schematic diagram showing the detection effect of a local anomaly factor algorithm of the prior art using a fixed neighborhood parameter.

[0025] Figure 3 is a schematic diagram showing the detection effect of a safety warning method in an embodiment of the present application using a dynamic neighborhood parameter. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, but not all the embodiments of the present application.

[0027] Figure 1 is a flowchart of a safety warning method for a powder alloy processing device according to an embodiment of the present application. As shown in Figure 1 the safety warning method for the powder alloy processing device includes steps S101 to S104, which are described in detail below.

[0028] S101, collecting running state data of the powder alloy processing device to generate a series of multi-dimensional feature vectors, and constructing a health state reference data set based on the multi-dimensional feature vectors.

[0029] In an embodiment, high-frequency three-axis vibration acceleration sensors and wide-band acoustic sensors can be deployed at key components of the powder alloy processing device, such as the atomizer nozzle base of the gas atomization powder production device, the motor bearing seat of the high-energy ball mill, the screen frame of the vibrating screen separator, and other core positions, to collect multi-channel time series signal data streams of the device in normal operation in real time at a high sampling rate, for example, 20 kHz.

[0030] In this optional embodiment, after collecting the original signals, signal preprocessing and windowing can be performed. The collected original signals are processed using a digital bandpass filter to filter out power frequency interference and high-frequency white noise that is not related to the state of the device. Then, a fixed time length T, for example, T=1 second, is used as the window size, and a certain overlap rate, for example, 50%, is used to slide on the time axis, so that the continuous signal stream is cut into a series of continuous signal segments.

[0031] Further, feature extraction is performed on each signal segment to generate a multi-dimensional feature vector. These feature indicators comprehensively cover time domain and frequency domain information, and can specifically include: time domain statistical features such as root mean square, kurtosis, skewness, etc., and frequency domain statistical features such as dominant frequency, spectral entropy, etc. Through this process, the continuous running state of the device is mapped to a series of data points in the feature space, i.e., a series of multi-dimensional feature vectors.

[0032] Finally, a health status reference dataset can be constructed based on the acquired multidimensional feature vectors. During the health confirmation phase after the equipment leaves the factory or after a major overhaul, long-term operating data under various known fault-free operating conditions, such as different loads and different ambient temperatures, can be collected and converted into multidimensional feature vectors. These vectors together constitute a health status reference dataset covering various normal operating modes of the equipment.

[0033] In this way, by collecting and extracting features from equipment status data, a multi-dimensional feature vector is generated for real-time monitoring and historical benchmarking, and a health status reference dataset is constructed as a benchmark for subsequent anomaly judgment.

[0034] S102, based on the fluctuation of the historical feature vectors corresponding to each reference data point in the health status reference dataset, label the reference data points with inherent stability weights that characterize their intrinsic stability.

[0035] In one embodiment, for each reference data point in the health status reference dataset The corresponding fluctuations in historical feature vectors, where historical feature vectors refer to the fluctuations in... This multidimensional feature vector generation generates M adjacent multidimensional feature vectors in the time series prior to the given time step; for example, M=10. Specifically, the geometric centroid of this local window composed of M vectors is calculated, then the Euclidean distance from each historical vector within the window to this centroid is calculated, and the average of these distances is used as the... Historical volatility.

[0036] Furthermore, an intrinsic stability weight representing the inherent stability of the reference data point can be generated based on historical volatility. This weight can be calculated using a decreasing exponential function, with the aim of mapping volatility to a weight. The intrinsic stability weight satisfies the following relationship:

[0037]

[0038] in, Reference data points The inherent stability weights have a range of (0,1]. Reference data points The historical volatility corresponding to the fluctuation of the historical eigenvector. and These represent the minimum and maximum values ​​of all historical volatility in the health status reference dataset, respectively. This is the weight decay coefficient, for example, a value of 3, used to adjust the rate at which the weight decreases as instability increases.

[0039] Thus, by assigning a weight to each point in the health dataset that reflects its reliability as a health sample, a foundation is laid for more accurate assessment of the local environment of the test point during subsequent online detection.

[0040] S103: Real-time acquisition of the data point to be tested; evaluation of the local sparsity of the data environment in which the data point to be tested is located based on the inherent stability weight of the neighboring data points of the data point to be tested; and application of risk penalty in combination with the instantaneous fluctuation of the data point to be tested itself, so as to determine the dynamic neighborhood parameters suitable for the data point to be tested.

[0041] In one embodiment, for the data points to be measured Find its health status reference dataset from the constructed dataset Its nearest neighbors constitute its initial neighborhood set. Exemplary The value is 50, which represents the size of the initial neighborhood.

[0042] Furthermore, the stability-weighted average distance between the data point to be tested and its initial neighborhood can be calculated. The stability-weighted average distance satisfies the following relationship:

[0043]

[0044] in, Data points to be tested The stability-weighted average distance, The initial neighborhood size, Data points to be tested The initial neighborhood, Data points to be tested With neighboring points European distance, Point for neighbors The inherent stability weights are used. By using the stability weights as the denominator, the distance is dynamically adjusted. The spatial distance of a neighboring point from an unstable period is amplified, thereby increasing the stability-weighted average distance and reflecting a sparser environment.

[0045] In this optional embodiment, each data point in the health status reference dataset can be traversed to calculate the stability weighted average distance corresponding to all data points, thereby obtaining the maximum and minimum values. When a new data point to be tested is transmitted in real time, the stability weighted average distance of the new data point to be tested can be normalized by the max-min normalization method. In this scheme, the normalized stability weighted average distance is used as the local sparsity of the data environment in which the data point to be tested is located.

[0046] Further, the instantaneous fluctuation degree of the current data point to be tested can be calculated, specifically, a preset M historical feature vectors before the data point to be tested are acquired, a geometric centroid of the historical feature vectors is calculated, and an average Euclidean distance from the historical feature vectors to the geometric centroid is calculated.

[0047] Then, the reference neighborhood parameter is calculated based on the obtained local sparsity, the reference neighborhood parameter satisfies the relationship:

[0048]

[0049] wherein, is a preset upper limit of the neighborhood parameter, is a sparsity index generated according to the local sparsity evaluation, is a sparsity response coefficient, and is calculated according to the instantaneous fluctuation degree A fluctuation penalty term is applied to the reference neighborhood parameter to obtain a final dynamic neighborhood parameter , and the dynamic neighborhood parameter satisfies the relationship:

[0050]

[0051] wherein, is the dynamic neighborhood parameter, is the reference neighborhood parameter determined based on the local sparsity evaluation result, is a preset lower limit of the neighborhood parameter, is the instantaneous fluctuation degree corresponding to the instantaneous fluctuation of the data point to be tested, is a fluctuation penalty coefficient, is a floor function.

[0052] Exemplarily, and are respectively 10 and 100, and are respectively 2 and 2.5, when the equipment fluctuates violently, the instantaneous fluctuation degree increases, which will cause the dynamic neighborhood parameter to be pressed to close to , so that a more sensitive analysis scale is taken.

[0053] In this way, through the hierarchical logical evaluation, an optimal neighborhood parameter is determined for each data point to be tested, and a key input is provided for the final accurate early warning.

[0054] S104, a local anomaly factor of the data point to be tested is calculated based on the dynamic neighborhood parameter, and whether the powder alloy processing equipment enters a sub-healthy state is judged according to the local anomaly factor, so as to realize the safety early warning of the powder alloy processing equipment.

[0055] In one embodiment, after obtaining the dynamic neighborhood parameter, a standard local outlier factor (LOF) algorithm process can be executed to calculate the local outlier factor score of the to-be-tested data point, and determine whether the powder alloy processing equipment enters a sub-healthy state according to the local outlier factor, specifically: comparing the calculated local outlier factor of the to-be-tested data point with a preset early warning threshold value, if the local outlier factor is greater than the early warning threshold value, it is determined that the powder alloy processing equipment enters a sub-healthy state and triggers an early warning signal, and sends an alarm information to the central monitoring system; if the local outlier factor is not greater than the early warning threshold value, it is determined that the powder alloy processing equipment is in a healthy state and no early warning is needed. In most industrial applications, a LOF value greater than 1.5 is usually considered a strong abnormal signal, and the early warning threshold value in the present scheme can be exemplarily set to 1.6.

[0056] As shown in Figure 2 and Figure 3 , they are respectively schematic detection effect diagrams of the local outlier factor algorithm of the prior art using a fixed neighborhood parameter and the safety early warning method of the embodiment of the present application using a dynamic neighborhood parameter. It can be seen that the prior art using a fixed neighborhood parameter value cannot identify the sub-healthy point represented by the red X, while in the embodiment of the present application, different dynamic neighborhood parameters are adaptively matched for data points in different regions and different states, a smaller dynamic neighborhood parameter value 12 is matched for the sub-healthy point, which is forced to be finely reviewed, so that its abnormal score is significantly amplified, and it is successfully detected.

[0057] In this way, by using dynamically adapted neighborhood parameters for final abnormality evaluation, early weak fault signs that are difficult to find by traditional methods can be accurately identified, and reliable monitoring and timely early warning of the sub-healthy state of the equipment are realized.

[0058] It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the protection scope of the present application patent should be subject to the appended claims.

Claims

1. A safety warning method for a powder alloy processing apparatus, characterized by, The method comprises the following steps: Collecting operation state data of a powder alloy processing device to generate a series of multi-dimensional feature vectors, and constructing a health state reference data set based on the multi-dimensional feature vectors; According to the fluctuation of the historical feature vector corresponding to each reference data point in the health state reference data set, the intrinsic stability weight of the reference data point is labeled to represent its inherent stability; Real-time acquisition of a to-be-tested data point, evaluation of the local sparsity of the data environment where the to-be-tested data point is located based on the intrinsic stability weight of the neighbor data point of the to-be-tested data point, and combination of the instantaneous fluctuation of the to-be-tested data point itself to impose a risk penalty, so as to determine the dynamic neighborhood parameter suitable for the to-be-tested data point; Based on the dynamic neighborhood parameter, the local anomaly factor of the to-be-tested data point is calculated, and whether the powder alloy processing device enters a sub-health state is judged according to the local anomaly factor, so as to realize the safety warning of the powder alloy processing device.

2. A safety warning method for a powder alloy processing apparatus according to claim 1, wherein The intrinsic stability weight satisfies the relationship: wherein, is an intrinsic stability weight of the reference data point is a history fluctuation degree corresponding to a fluctuation situation of a history feature vector of the reference data point and are respectively a minimum value and a maximum value of all history fluctuation degrees in the health status reference data set, is a preset weight attenuation coefficient.​​ 3. The safety warning method for powder alloy processing equipment according to claim 1, wherein, The evaluation process of the local sparsity is to determine the initial neighborhood of the to-be-tested data point, and to calculate the stability weighted average distance between the to-be-tested data point and each neighbor point in the initial neighborhood, which satisfies the relationship: wherein, is the stability weighted average distance of the data point to be tested, is the size of the initial neighborhood, is the initial neighborhood of the data point to be tested, is the Euclidean distance of the data point to be tested to the neighbor point , is the intrinsic stability weight of the neighbor point .

4. The safety warning method for powder alloy processing equipment according to claim 1, wherein, The dynamic neighborhood parameter satisfies the relationship: wherein, is a dynamic neighborhood parameter, is a reference neighborhood parameter determined based on a local sparsity evaluation result, is a preset lower limit of the neighborhood parameter, is an instantaneous fluctuation degree corresponding to an instantaneous fluctuation condition of the data point to be measured, is a fluctuation penalty coefficient, is a floor function.

5. The safety warning method for powder alloy processing equipment according to claim 1, wherein, The multi-dimensional feature vector comprises time domain statistical features and frequency domain statistical features; the time domain statistical features comprise root mean square, kurtosis and skewness; and the frequency domain statistical features comprise main frequency and spectral entropy.

6. A safety warning method for a powder alloy processing apparatus according to claim 2, wherein The calculation method of the historical fluctuation degree is to obtain a preset number of historical feature vectors before the reference data point, calculate the geometric centroid of the historical feature vectors, and calculate the average Euclidean distance from the historical feature vectors to the geometric centroid.

7. A safety warning method for a powder alloy processing apparatus according to claim 4, wherein The calculation method of the instantaneous fluctuation degree is to obtain a preset number of historical feature vectors before the to-be-tested data point, calculate the geometric centroid of the historical feature vectors, and calculate the average Euclidean distance from the historical feature vectors to the geometric centroid.

8. A safety warning method for a powder alloy processing apparatus according to claim 4, wherein The reference neighborhood parameter satisfies the relationship: wherein, is a predetermined upper limit of the neighborhood parameter, is a sparsity index generated according to a local sparsity assessment, is a sparsity response coefficient.

9. The safety warning method for powder alloy processing equipment according to claim 1, wherein, The calculation of the local anomaly factor of the to-be-tested data point based on the dynamic neighborhood parameter is specifically to calculate the local anomaly factor score of the to-be-tested data point by using a local anomaly factor algorithm.

10. The safety warning method for powder alloy processing equipment according to claim 1, wherein, The judgment of whether the powder alloy processing device enters a sub-health state according to the local anomaly factor comprises: Comparing the calculated local anomaly factor of the to-be-tested data point with a preset warning threshold; If the local anomaly factor is greater than the warning threshold, it is determined that the powder alloy processing device enters a sub-health state and triggers a warning signal; If the local anomaly factor is not greater than the warning threshold, it is determined that the state of the powder alloy processing device is healthy.

Images