Method for monitoring operating parameters of plant equipment based on intelligent sensors

Abstract

The application relates to the technical field of intelligent manufacturing, in particular to a workshop equipment operation parameter monitoring method based on an intelligent sensor, which comprises the following steps: acquiring a workshop equipment operation parameter sampling sequence collected by a plurality of sensors, calculating normalized first-order difference gradients of the sampling sequences of the sensors, and constructing a multi-sensor joint boundary activation score to determine a working condition switching boundary event; taking the working condition switching boundary event as a time segmentation anchor point to cut an adaptive window, and associating a working condition category identifier; extracting time domain characteristic components and frequency domain characteristic components in the adaptive window, and splicing the time domain characteristic components and the frequency domain characteristic components into a window-level feature vector; acquiring a feature baseline corresponding to the working condition category identifier, calculating Mahalanobis distances of the window-level feature vector relative to the feature baseline, and determining that a device state is abnormal and triggering an alarm when the Mahalanobis distances are greater than an alarm determination threshold. The application eliminates cross-working condition data aliasing, and improves the detectability of early abnormal signs at the moment of working condition switching.

Description

Technical Field

[0001] This application relates to the field of intelligent manufacturing technology, specifically to a method for monitoring the operating parameters of workshop equipment based on intelligent sensors. Background Technology

[0002] In the field of smart manufacturing, the health status of workshop equipment directly determines the utilization rate of production lines and product yield. Core equipment such as CNC machine tools, industrial robots, and conveyor belt drive motors continuously generate multi-dimensional operating parameters such as vibration, temperature, current, and pressure during processing. These parameters carry crucial information about the mechanical status of the equipment. With the rapid decline in the cost of smart sensors and the improvement of edge computing capabilities, real-time equipment monitoring based on sensor networks has become a standard infrastructure in smart workshops.

[0003] Existing monitoring methods generally adopt a processing flow of fixed time window plus statistical feature extraction plus threshold alarm. The system segments the original sampling sequence of the sensor with a fixed length, calculates statistical features such as mean, variance, and peak factor in each window, and then compares the extracted features with the preset threshold. If the threshold is exceeded, an alarm is triggered.

[0004] However, taking CNC machine tools as an example, a complete machining cycle includes multiple operating conditions such as no-load positioning, tool entry, stable cutting, and tool retraction and change, with significant differences in vibration amplitude and current load in each segment. When the operating condition switching time falls exactly in the middle of a fixed time window, the window simultaneously mixes data from two segments with different characteristics before and after the switch: if there is an abnormal pulse caused by an early fault after the switch, its duration often only accounts for a small portion of the total window length. The statistical averaging effect of the normal operating condition data in the first half will significantly suppress the characteristics of this pulse, causing the calculated window-level characteristic value to still fall within the normal threshold band, forming a systemic anomaly detection blind zone. This blind zone is precisely the stage where the fault signal is weakest and most in need of being captured in the early signs of equipment degradation, but it is ignored by the system due to the overlapping of operating condition data across the window, and is only perceived when the fault evolves into a sudden shutdown. Summary of the Invention

[0005] To address the existing technical problem of detection blind spots caused by the homogenization and dilution of early abnormal features due to cross-condition data aliasing during equipment condition switching within a fixed time window, this application provides a method for monitoring workshop equipment operating parameters based on intelligent sensors.

[0006] In a first aspect, this application provides a method for monitoring workshop equipment operating parameters based on intelligent sensors, comprising: acquiring sampling sequences of workshop equipment operating parameters collected by multiple sensors; calculating the normalized first-order differential gradient of each sensor sampling sequence; constructing a multi-sensor joint boundary activation score based on the normalized first-order differential gradient; determining a working condition switching boundary event when the multi-sensor joint boundary activation score exceeds a multi-sensor activation ratio threshold; using the working condition switching boundary event as a time segmentation anchor point to segment an adaptive window between two adjacent working condition switching boundary events; and using the processing program segment number issued by the equipment side as the working condition category identifier associated with the adaptive window; extracting time-domain feature components and frequency-domain feature components from the sampling sequences of each sensor within the adaptive window; concatenating the extracted time-domain feature components and frequency-domain feature components into a window-level feature vector corresponding to the adaptive window; acquiring a feature baseline constructed from the historical normal feature sample set corresponding to the working condition category identifier; calculating the Mahalanobis distance between the window-level feature vector and the feature baseline; and determining an abnormal equipment state within the adaptive window and triggering an alarm when the Mahalanobis distance is greater than an alarm determination threshold.

[0007] By using multiple sensors to jointly determine the boundary of operating condition switching and segment the adaptive window, cross-operating condition aliasing is eliminated from the source of data segmentation, ensuring that the data in each analysis window belongs to the same operating condition segment. This avoids data pollution introduced by random phase deviation in fixed windows and significantly improves the detectability of early abnormal signs at the moment of operating condition switching.

[0008] Preferably, the calculation of the normalized first-order differential gradient of the sampling sequences of each sensor includes: calculating the absolute difference between the current sampling point and the previous sampling point of each sensor sampling sequence; obtaining the standard deviation of the sampling sequences of each sensor in the recent stable operation period; dividing the absolute difference by the standard deviation to finally obtain the normalized first-order differential gradient.

[0009] By calculating the normalized first-order differential gradient, the sensor gradients of different dimensions are normalized to a dimensionless space that can be compared with each other, eliminating the comparability barrier between channels of different dimensions and preventing channels with larger magnitudes from dominating the judgment results.

[0010] Preferably, the step of constructing the multi-sensor joint boundary activation score based on the normalized first-order differential gradient includes: comparing the normalized first-order differential gradient of each sensor with the corresponding adaptive trigger threshold; counting the number of sensor paths whose normalized first-order differential gradient is greater than the adaptive trigger threshold; and dividing the number of sensor paths by the total number of sensors to finally obtain the multi-sensor joint boundary activation score.

[0011] Preferably, the step of segmenting to obtain an adaptive window between two adjacent working condition switching boundary events includes: calculating the rotation cycle corresponding to the rated speed of the workshop equipment; using the duration of two rotation cycles as the required shortest window duration; determining whether the time interval between two adjacent working condition switching boundary events is less than the required shortest window duration; if it is less, merging the window corresponding to the current time interval with the subsequent windows in sequence until the cumulative duration after merging reaches the required shortest window duration, and finally obtaining the adaptive window.

[0012] By setting the required minimum window duration and merging windows, it is ensured that the window contains at least a complete rotation cycle, so that the frequency domain feature extraction is physically harmonic distinguishable, avoiding feature extraction failure caused by dense process switching.

[0013] Preferably, the step of extracting time-domain feature components from the sampling sequences of each sensor within the adaptive window includes: calculating the root mean square value of each sensor sampling sequence within the adaptive window as an energy level feature; calculating the ratio of the peak value to the root mean square value of each sensor sampling sequence within the adaptive window as a peak factor feature; calculating the ratio of the root mean square value to the absolute mean as a waveform exponential feature; and combining the energy level feature, the peak factor feature, and the waveform exponential feature to finally obtain the time-domain feature components.

[0014] Preferably, the step of extracting frequency domain feature components from the sampling sequences of each sensor within the adaptive window includes: performing a fast Fourier transform on the sampling sequences of each sensor within the adaptive window to obtain spectral data; calculating the rotational frequency based on the real-time rotational speed signal of the workshop equipment; extracting the rotational frequency and the amplitude component at a preset order harmonic of the rotational frequency from the spectral data; calculating the ratio of narrowband energy centered on the rotational frequency and having a preset bandwidth to the total spectral energy; and combining the amplitude component with the ratio to finally obtain the frequency domain feature components.

[0015] Preferably, the step of obtaining the feature baseline constructed from the historical normal feature sample set corresponding to the working condition category identifier includes: extracting the mean vector of the historical normal feature sample set; calculating the covariance matrix of the historical normal feature sample set; and combining the mean vector and the covariance matrix to finally obtain the feature baseline.

[0016] Preferably, calculating the Mahalanobis distance between the window-level feature vector and the feature baseline includes: calculating the difference vector between the window-level feature vector and the mean vector; obtaining the inverse matrix of the covariance matrix; and performing a square root operation based on the transpose of the difference vector, the inverse matrix, and the product of the difference vector to finally obtain the Mahalanobis distance.

[0017] Preferably, the step of determining that the device status within the adaptive window is abnormal and triggering an alarm includes: determining that the device status within the adaptive window is abnormal and triggering a hierarchical alarm when the Mahalanobis distance is greater than the alarm determination threshold; calculating the dimension-wise deviation vector between the window-level feature vector and the mean vector; sorting the absolute values ​​of the deviations of each dimension in the dimension-wise deviation vector in descending order; extracting the top three feature dimensions with the largest absolute deviation values ​​as anomaly tracing indicators, and finally completing the determination and alarm of the device status abnormality.

[0018] By extracting the top three feature dimensions with the largest absolute deviation as anomaly tracing indicators, the abstract distance calculation results are mapped to specific fault types, assisting maintenance personnel in quickly determining fault types and improving the practicality of the monitoring system.

[0019] Preferably, the step of obtaining the feature baseline constructed from the historical normal feature sample set corresponding to the working condition category identifier further includes: counting the number of windows that continuously trigger alarms; when the number of windows exceeds a preset multiple of the average number of working condition cycles for the same working condition category during the calibration phase, including the feature samples within the continuous alarm segment into the baseline candidate pool; and re-estimating the mean vector and covariance matrix using the feature distribution in the baseline candidate pool to finally obtain the updated feature baseline.

[0020] This application establishes a feature baseline independently based on operating condition classification, eliminating false alarm interference caused by differences in feature distribution across operating conditions, and making the anomaly judgment results statistically clear and probabilistically interpretable. At the same time, this application has out-of-the-box adaptive capability in different equipment types and workshop environments, reducing the engineering deployment threshold. With the help of the dual-path baseline update mechanism, it takes into account the distinction between equipment natural aging drift and occasional abnormal alarms, so that the monitoring system maintains continuous and effective discrimination capability in long-term operation. Attached Figure Description

[0021] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein: Figure 1 This is a flowchart illustrating a method for monitoring workshop equipment operating parameters based on intelligent sensors according to the present invention.

[0022] Figure 2 This is a two-dimensional distribution comparison chart of the detection rate and false alarm rate of the existing method and the method of this application under various working conditions in the embodiments of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] This invention discloses a method for monitoring workshop equipment operating parameters based on intelligent sensors, referring to... Figure 1 This includes steps S1-S4: S1. Obtain the sampling sequence to determine the boundary event of the operating condition switching.

[0026] In an optional embodiment, the system is deployed. The road sensor collects data on the workshop equipment in real time. Road sensors in The original sampled value at time t is denoted as This involves acquiring sampling sequences of workshop equipment operating parameters from multiple sensors. Due to significant differences in dimensions and magnitudes among different sensor channels, the unit for spindle current is amperes, and the unit for vibration acceleration is... Since the absolute gradient values ​​of the two are incomparable in magnitude, if a joint decision is made by directly performing point-by-point differences on the original sampled values, the channel with the larger magnitude will dominate the judgment result, and the abrupt change information of the channel with the smaller magnitude will be submerged. Therefore, at each sampling point, the absolute difference between the current sampling point and the previous sampling point of each sensor's sampling sequence is calculated, and the normalized first-order difference gradient is calculated for each sensor separately:

[0027] in For the first The standard deviation of the sampling sequence of the road sensor in the recent stable operation period is specifically calculated from the measured sampling data in the most recent complete operating condition window. The absolute difference is divided by the standard deviation to obtain the normalized first-order differential gradient, which is used to normalize the sensor gradients of different dimensions to a dimensionless space that can be compared with each other. It is updated on a rolling basis after each complete operating condition window ends to continuously track the long-term drift of the sensor noise floor level.

[0028] Furthermore, based on the normalized first-order difference gradients of each path, a multi-sensor joint boundary activation score is constructed. :

[0029] in For the first The activation determination item of the road sensor, when the first Path normalized gradient Exceeding the adaptive trigger threshold of this path hour Set the value to 1, otherwise set it to 0. Compare the normalized first-order differential gradient of each sensor with its corresponding adaptive trigger threshold. Compare the data, count the number of sensors whose normalized first-order differential gradient is greater than the adaptive trigger threshold, and divide the number of sensors by the total number of sensors. Finally, the joint boundary activation score of the multi-sensor system is obtained. Threshold for each path. The determination method is as follows: during the calibration phase after the equipment completes the initialization calibration phase, data from at least three complete operating condition cycles must be collected—statistics of the first cycle are then calculated. The 95th percentile of the gradient value of the road sensor within the stable range of each operating condition is taken as the 95th percentile value. The initial value represents the upper bound of the fluctuation under normal operating conditions. Subsequently, at the end of each complete operating condition window, the newly accumulated stable segment data is used to... Rolling corrections are performed to make the threshold continuously adapt to changes in the sensor's noise floor level.

[0030] Then, when Exceeding the multi-sensor activation ratio threshold At that time, the judgment Constantly receive boundary events for operating condition switching . The determination is based on the statistics of known switching events from the calibration phase: using the known working condition switching times marked by the machining program segment numbers issued by the CNC system or PLC as reference true values, the proportional distribution of the actual synchronous triggering of each sensor when each real switching event occurs is statistically analyzed, and the 5th percentile of this distribution is taken as the true value. The target value is to ensure that at least 95% of known real switching events can be successfully detected, thereby achieving a balance between the false detection rate and the false trigger rate. The data is continuously updated with each complete cycle of known switching events, using newly accumulated data.

[0031] In actual workshop electromagnetic interference environments, gradient spikes caused by local interference in a single sensor channel can occur at any time. The probability of multiple sensors simultaneously exhibiting gradient abrupt changes is far lower than the probability of sporadic interference in a single channel. Multi-channel joint activation determination can effectively suppress the false trigger rate to an acceptable level. It should be noted that there are many other methods in this field for achieving normalized gradient calculation. Those skilled in the art can choose different normalization benchmarks based on the actual sensor layout. For example, using the interquartile range of historical data instead of the standard deviation as the normalization denominator does not affect the core logic of this solution.

[0032] Thus, by using a multi-sensor normalized gradient multi-path joint activation ratio determination method to achieve adaptive positioning of the working condition switching boundary, the comparability barrier between channels with different dimensions is eliminated, and the multi-path synchronous verification effectively suppresses misjudgment caused by local interference in a single channel, enabling the boundary event sequence to have high time positioning accuracy and providing a reliable time anchor for subsequent homogeneous window segmentation.

[0033] S2. Divide the adaptive window using boundary events as anchor points.

[0034] In an optional embodiment, the detected operating condition switching boundary event sequence is... Using these points as time segmentation anchors, an adaptive window is formed by dividing the time interval between two adjacent working condition switching boundary events. Since the start and end times of each window are strictly aligned with the moment of operating condition switching, the sampled data within the window must physically belong to the same operating condition segment, fundamentally eliminating the problem of cross-operating condition data aliasing caused by random phase deviation between the dividing point and the moment of operating condition switching in a fixed time window.

[0035] However, in some processing procedures, the switching between steps is extremely frequent, and the interval between adjacent boundary events may be as short as tens of sampling points, resulting in a window that is too short to support meaningful frequency domain feature extraction. Since in frequency domain analysis, when performing a Fast Fourier Transform on the sampled sequence within the window, the spectral resolution is... ,in The window duration is used to ensure that adjacent frequency harmonics are distinguishable in the spectrum; that is, the interval between adjacent harmonics is the frequency. The distinguishable requirement is to meet Therefore, the required minimum window duration can be derived:

[0036] in , Given the equipment's rated speed in rpm, calculate the rotational cycle corresponding to the rated speed of the workshop equipment, and use the duration of two rotational cycles as the required shortest window duration. This lower bound ensures that the window contains at least two complete rotation cycles, making the frequency domain feature extraction physically harmonic distinguishable. It determines whether the time interval between two adjacent operating condition switching boundary events is less than the required shortest window duration. If the interval between any two adjacent boundary events is less than... Then, the window corresponding to the current time interval will be merged with its subsequent windows in sequence until the cumulative duration of the merged windows reaches the required minimum window duration. Finally, an adaptive window is obtained. The merged windows are then categorized using the initial operating condition label in subsequent steps.

[0037] Furthermore, the processing program segment numbers issued by the equipment side are used as adaptive windows. Related operating condition category identifier In manufacturing data access systems equipped with protocol conversion modules, program segment numbers can be synchronized to the monitoring system in real time via standard industrial protocols, aligning with sensor timestamps without requiring additional manual annotation. After segmentation, the original continuous sensor data stream is transformed into a series of time segments with clearly defined operating condition labels, and the data purity within each window is guaranteed from the source by a boundary anchoring mechanism.

[0038] In this way, the adaptive window dynamically constructed with the working condition switching boundary event as the segmentation anchor point ensures the working condition consistency of the data within each analysis unit, eliminates the data pollution introduced by random phase deviation in the fixed time window, and provides a guarantee for the physical interpretability of the subsequent feature extraction stage.

[0039] S3. Extract the time-domain feature components and frequency-domain feature components within the adaptive window and concatenate them into a window-level feature vector.

[0040] In an optional embodiment, for each adaptive window Within the adaptive window, time-domain and frequency-domain feature components are extracted from the sampling sequences of each sensor, and the two components describe the operating status of the device from different perspectives.

[0041] In the time domain dimension, for adaptive windows Inner Three feature components reflecting different physical meanings are extracted from the sampling sequences of the road sensors. The root mean square value of the sampling sequences of each road sensor within the adaptive window is calculated. As an energy level characteristic, it reflects the energy level of the operating condition and is a fundamental quantity for judging load intensity; the ratio of the peak value to the root mean square value of each sensor sampling sequence within the adaptive window is used as the peak factor characteristic. Under normal rolling bearing operating conditions, the peak factor remains at a low and stable level. When pitting or spalling-type impact failures occur, the peak factor shows a significant jump. The larger the value, the more pronounced the deviation of the transient impact amplitude in the sampling sequence relative to the overall energy level, and the higher the suspicion of an impact failure. The ratio of the root mean square value to the absolute mean is calculated as the waveform exponential feature. It is more sensitive to waveform distortion caused by early distributed wear. The closer this value is to the theoretical value of a sine wave, the more regular the waveform shape is, and the lower the degree of waveform distortion caused by early wear.

[0042] By combining energy level features, peak factor features, and waveform exponential features, the time-domain feature components are finally obtained. These three time-domain features characterize the time-domain state of the device from three orthogonal dimensions: energy intensity, impact degree, and waveform shape, with no information redundancy between them.

[0043] In the frequency domain, a fast Fourier transform is performed on the sampling sequences of each sensor within the adaptive window to obtain spectral data. Then, the rotational frequency is calculated based on the real-time rotational speed signal of the workshop equipment. ,in , The current rotational speed is expressed in rpm, ensuring that the center frequency of the frequency domain feature extraction is synchronized with the actual operating status of the equipment.

[0044] Extracting frequency conversion from spectrum data In addition, the amplitude components at the preset order harmonics of the frequency conversion can be extracted. The preset order is determined based on the typical fault frequency characteristics of the equipment type. For example, the amplitude components at the second and third harmonics of the frequency conversion can be extracted. , , The frequency harmonic energy reflects the frequency domain characteristics of rotational faults such as rotor imbalance and misalignment. Simultaneously, the ratio of narrowband energy centered on the frequency and with a preset bandwidth to the total spectral energy is calculated. The preset bandwidth is determined based on the characteristic frequency spacing corresponding to the bearing model of the equipment. For example, the preset bandwidth can be taken as... This ratio reflects the degree of energy concentration near the rotational frequency. A higher value indicates that the energy is more concentrated near the rotational frequency, and the more regular the rotor's dynamic characteristics are. When resonance or instability occurs in the shaft system... An abnormal drop will occur, indicating that energy has diffused to other frequency bands. Combining the amplitude component with the ratio, the frequency domain characteristic components are finally obtained.

[0045] The extracted time-domain feature components and frequency-domain feature components are concatenated to form the window-level feature vector of the corresponding adaptive window. Each sensor contributes three time-domain components and a preset number of frequency-domain components of corresponding order. This vector fully encapsulates the device's performance under the current operating conditions. The multidimensional state information within is used as input for subsequent anomaly assessment.

[0046] Thus, the multidimensional time-domain and frequency-domain feature vectors extracted within the adaptive window can accurately reflect the true equipment status of the current operating condition due to the purity of the window data. The slight feature deviations caused by early faults are no longer masked by the data mixed across operating conditions, laying a reliable feature input foundation for subsequent abnormal deviation assessment based on Mahalanobis distance.

[0047] S4. Calculate the Mahalanobis distance to determine if the equipment is in an abnormal state and trigger an alarm.

[0048] In an optional embodiment, for each type of operating condition Maintain an independent feature baseline by acquiring a set of historical normal feature samples corresponding to the working condition category identifier to construct the feature baseline. Let the working condition be... Historical normal feature sample set is Extract the mean vector of the historical normal feature sample set. Calculate the covariance matrix of the historical normal feature sample set. The mean vector and covariance matrix are combined to obtain the characteristic baseline. The normal operating characteristic distribution of equipment under different working conditions essentially belongs to different statistical populations. The root mean square vibration of the cutting condition is much higher than that of the no-load condition. Directly comparing characteristic vectors across working conditions would generate a large number of false alarms unrelated to the differences in working conditions themselves. Establishing independent baselines for each working condition ensures that each anomaly determination is made within the historical distribution of the same working condition, eliminating false alarm interference introduced by differences in working conditions at the source.

[0049] For the current evaluation of the first window-level feature vectors (Working condition category is) ), calculate the Mahalanobis distance of the window-level feature vectors relative to the feature baseline:

[0050] Specifically, it calculates the difference vector between the window-level feature vector and the mean vector. Obtain the inverse of the covariance matrix. The Mahalanobis distance is obtained by taking the square root of the transpose, inverse matrix, and product of the difference vectors. Mahalanobis distance passes through The system automatically eliminates the influence of differences in the dimensions of different features, while also considering the correlation structure between features, thus avoiding the problem of redundant feature duplication and amplification. Under the assumption that the feature vectors follow a multivariate normal distribution, the squared Mahalanobis distance follows a distribution with [degrees of freedom missing]. ( The chi-square distribution (where the dimension of the window-level feature vector is 1) is used, and the Mahalanobis distance value corresponding to the 97.5 percentile of the chi-square distribution is taken as the alarm judgment threshold. This means that under normal operating conditions, 97.5% of the sample points have a Mahalanobis distance below this threshold, thus controlling the false alarm rate to a statistically acceptable level.

[0051] Furthermore, when the Mahalanobis distance is greater than the alarm determination threshold... When the device status within the adaptive window is abnormal, a tiered alarm is triggered. To further pinpoint the root cause of the anomaly, the dimension-wise deviation vector between the window-level feature vector and the mean vector is calculated. The absolute values ​​of the deviations in each dimension of the dimension-wise deviation vector are sorted in descending order, and the three feature dimensions with the largest absolute deviation values ​​are extracted as anomaly tracing indicators. This process ultimately completes the determination and alarm of the device status anomaly, assisting maintenance personnel in quickly identifying the fault type: if the deviation in the vibration peak factor dimension is the most prominent, it points to an impact fault; if the deviation in the frequency second harmonic energy dimension is the largest, it points to a misalignment fault.

[0052] Next, the baseline and The system uses a sliding update method for continuous updates. The normal update path is as follows: whenever the number of newly accumulated normal judgment window samples makes the current working condition... mean vector When the width of the 95% confidence interval for each dimension narrows to a preset narrowing ratio corresponding to the width when the initial baseline was established, an incremental update is triggered. For example, the preset narrowing ratio can be 1 / 2.

[0053] Simultaneously, a forced update bypass is configured: the number of consecutive alarm trigger windows is counted, and during the calibration phase, the number of operating condition windows appearing in a complete processing cycle for the same operating condition category is counted. The average value of multiple processing cycles is used as the average number of operating condition cycles for that operating condition category and recorded for later use. When the number of consecutive alarm windows exceeds a preset multiple of the average number of operating condition cycles, the system determines that the equipment has entered a new steady state rather than an occasional anomaly. The feature samples within the consecutive alarm segment are included in the baseline candidate pool, and the mean vector and covariance matrix are re-estimated using the feature distribution in the baseline candidate pool. Finally, the updated feature baseline is obtained, and an equipment state drift warning is reported. After confirmation by the operation and maintenance personnel, it is decided whether to accept the new baseline, thus preventing the one-way lock-up problem of continuous false alarms due to baseline freezing after the equipment enters a slow deterioration state.

[0054] like Figure 2 The figure shows a two-dimensional distribution comparison of the detection rate and false alarm rate of the existing method and the method of this application under various operating conditions in the embodiments of the present invention. It can be seen that the sample points of the existing method are generally concentrated in the region of high false alarm rate and low detection rate, while the sample points of the method of this application are generally migrated to the region of low false alarm rate and high detection rate. There is no overlap between the two sets of confidence ellipses, indicating that the method of this application achieves better detection performance and lower false alarm level under various operating conditions.

[0055] Thus, the Mahalanobis distance anomaly assessment based on the operating condition classification baseline not only provides a clear probabilistic basis for each alarm at the statistical level, but also maps the abstract distance calculation results to specific fault types through dimension-by-dimensional deviation tracing. At the same time, the dual-path baseline update mechanism takes into account the distinction between equipment natural aging drift and occasional abnormal alarms, enabling the monitoring system to maintain a continuous and effective discrimination capability during long-term operation.

Claims

1. A method for monitoring workshop equipment operating parameters based on intelligent sensors, characterized in that, include: The process involves acquiring sampling sequences of workshop equipment operating parameters collected by multiple sensors, calculating the normalized first-order differential gradient of each sensor sampling sequence, and constructing a multi-sensor joint boundary activation score based on the normalized first-order differential gradient. When the multi-sensor joint boundary activation score exceeds the multi-sensor activation ratio threshold, a working condition switching boundary event is determined. Using the working condition switching boundary event as the time segmentation anchor point, an adaptive window is obtained between two adjacent working condition switching boundary events, and the working condition category identifier associated with the adaptive window is combined with the processing program segment number issued by the equipment side. Within the adaptive window, time-domain feature components and frequency-domain feature components are extracted from the sampling sequences of each sensor, and the extracted time-domain feature components and frequency-domain feature components are concatenated to form a window-level feature vector corresponding to the adaptive window. Obtain the feature baseline constructed from the historical normal feature sample set corresponding to the working condition category identifier, calculate the Mahalanobis distance of the window-level feature vector relative to the feature baseline, and when the Mahalanobis distance is greater than the alarm judgment threshold, determine that the device status within the adaptive window is abnormal and trigger an alarm.

2. The method for monitoring workshop equipment operating parameters based on intelligent sensors according to claim 1, characterized in that, The calculation of the normalized first-order differential gradient of the sampling sequences of each sensor includes: Calculate the absolute difference between the current sampling point and the previous sampling point for each sensor sampling sequence; Obtain the standard deviation of the sampling sequences of each sensor during the recent stable operating period; Dividing the absolute difference by the standard deviation yields the normalized first-order difference gradient.

3. The method for monitoring workshop equipment operating parameters based on intelligent sensors according to claim 1, characterized in that, The construction of the multi-sensor joint boundary activation score based on the normalized first-order difference gradient includes: The normalized first-order differential gradient of each sensor is compared with the corresponding adaptive trigger threshold. Count the number of sensor channels whose normalized first-order differential gradient is greater than the adaptive trigger threshold; Dividing the number of sensor channels by the total number of sensor channels yields the multi-sensor joint boundary activation score.

4. The method for monitoring workshop equipment operating parameters based on intelligent sensors according to claim 1, characterized in that, The segmentation obtains an adaptive window between two adjacent working condition switching boundary events, including: Calculate the rotation period corresponding to the rated speed of the workshop equipment; The duration of the two rotation cycles is taken as the required shortest window duration; Determine whether the time interval between two adjacent working condition switching boundary events is less than the required shortest window duration; If the time interval is less than the required minimum window duration, the window corresponding to the current time interval is merged with the subsequent windows in sequence until the cumulative duration of the merged window reaches the required minimum window duration, and finally the adaptive window is obtained.

5. The method for monitoring workshop equipment operating parameters based on intelligent sensors according to claim 1, characterized in that, The step of extracting temporal feature components from the sampling sequences of each sensor within the adaptive window includes: The root mean square value of the sampling sequence of each sensor within the adaptive window is calculated as the energy level feature; The ratio of the peak value to the root mean square value of each sensor sampling sequence within the adaptive window is calculated as the peak factor feature. The ratio of the root mean square value to the absolute mean is calculated as a waveform exponential feature; The energy level feature, the peak factor feature, and the waveform exponential feature are combined to obtain the time-domain feature component.

6. The method for monitoring workshop equipment operating parameters based on intelligent sensors according to claim 1, characterized in that, The step of extracting frequency domain feature components from the sampling sequences of each sensor within the adaptive window includes: Perform a Fast Fourier Transform on the sampling sequences of each sensor within the adaptive window to obtain spectral data; The rotational frequency is calculated based on the real-time rotational speed signal of the workshop equipment. Extract the frequency conversion and the amplitude component at a preset order harmonic of the frequency conversion from the spectrum data; Calculate the ratio of narrowband energy with a preset bandwidth centered on the frequency to the total spectrum energy; The amplitude component is combined with the ratio to obtain the frequency domain feature component.

7. The method for monitoring workshop equipment operating parameters based on intelligent sensors according to claim 1, characterized in that, The step of obtaining the feature baseline constructed from the historical normal feature sample set corresponding to the working condition category identifier includes: Extract the mean vector of the historical normal feature sample set; Calculate the covariance matrix of the historical normal feature sample set; The mean vector and the covariance matrix are combined to obtain the feature baseline.

8. The method for monitoring workshop equipment operating parameters based on intelligent sensors according to claim 1, characterized in that, The calculation of the Mahalanobis distance between the window-level feature vector and the feature baseline includes: Calculate the difference vector between the window-level feature vector and the mean vector; Obtain the inverse of the covariance matrix; The Mahalanobis distance is obtained by performing a square root operation on the transpose of the difference vector, the inverse matrix, and the product of the difference vector.

9. The method for monitoring workshop equipment operating parameters based on intelligent sensors according to claim 1, characterized in that, The step of determining that the device status within the adaptive window is abnormal and triggering an alarm includes: When the Mahalanobis distance is greater than the alarm determination threshold, the device status within the adaptive window is determined to be abnormal and a graded alarm is triggered. Calculate the dimension-wise deviation vector between the window-level feature vector and the mean vector; Arrange the absolute values ​​of the deviations in each dimension of the dimension-by-dimensional deviation vector in descending order; The top three feature dimensions with the largest absolute deviation values ​​are extracted as anomaly tracing indicators, and the determination and alarm of the abnormal device status are finally completed.

10. The method for monitoring workshop equipment operating parameters based on intelligent sensors according to claim 1, characterized in that, The step of obtaining the feature baseline constructed from the historical normal feature sample set corresponding to the working condition category identifier further includes: Count the number of windows that trigger alarms consecutively; When the number of windows exceeds a preset multiple of the average number of operating cycles for the same operating condition category during the calibration phase, the feature samples within the continuous alarm segment will be included in the baseline candidate pool. The mean vector and covariance matrix are re-estimated using the feature distribution in the baseline candidate pool, and the updated feature baseline is finally obtained.

Images