Rotary machine pump vibration data real-time analysis and early warning system based on edge side computing

Abstract

The present application relates to transmission link early warning technical field, disclose a real-time analysis and early warning system for rotating machine pump vibration data based on edge side calculation, include: capture the physical layer error code time sequence of transmission link and build the differential error code time sequence vector which removes the influence of absolute time drift;A dynamic rotating virtual electric vector space is constructed in the edge computing node, and the gradient ascent algorithm is used to optimize the target functional which measures the phase clustering strength, and the virtual space rotating frequency is adaptively adjusted to lock the best virtual frequency which makes the error code projection most concentrated;Based on the best frequency, the probability density distribution of error code phase is reconstructed by using kernel density estimation, and the generalized Renyi entropy is calculated as the characteristic of transmission error code phase clustering dispersion degree;By comparing the entropy characteristic with the dynamic threshold based on the environmental noise base, the transmission link abnormality is determined.

Description

Technical Field

[0001] This invention relates to the field of transmission link early warning technology, and more specifically, to a real-time analysis and early warning system for vibration data of rotating pumps based on edge-side computing. Background Technology

[0002] With the development of Industrial Internet of Things (IIoT) technology, online monitoring and fault diagnosis for critical rotating machinery such as pumps and turbines have shifted from traditional localized inspections to an architecture based on edge sensing and cloud analytics. In this architecture, edge computing gateways deployed next to the pumps and turbines are responsible for the acquisition, preprocessing, and uploading of high-frequency vibration data. Because vibration data is characterized by high throughput and strong time sensitivity, the stability of the data transmission link directly determines the effectiveness of remote diagnosis. Existing transmission link monitoring technologies typically follow the evaluation system of general IT networks, focusing primarily on macro-level indicators such as network bandwidth utilization, packet loss rate, and average round-trip time (RTT), using heartbeat detection (Ping) or Link Layer Discovery Protocol (LLDP) to determine network connectivity and congestion.

[0003] However, in actual industrial environments of rotating pumps and motors, the edge-side environment is extremely complex, especially with the widespread adoption of variable frequency drives (VFDs) introducing strong electromagnetic interference. Existing transmission link early warning methods have significant blind spots: current technologies are designed based on data transmission packet loss or delays caused by logical or external factors such as network traffic congestion or weak signal coverage, resulting in random bit error distribution (Poisson or uniform distribution). However, in pump and motor scenarios using space vector pulse width modulation (SVPWM) control, the high-frequency common-mode voltage and transient pulses (dV / dt) generated by the inverter can couple to the transmission cables through parasitic capacitance, causing a special anomaly in the transmission link. This anomaly manifests as the physical layer bit error rate (BER) no longer being random, but rather as periodic pulse errors deeply modulated by the pump and motor's electrical operating frequency. Existing technologies cannot identify the temporal rhythm characteristics of these errors, often misjudging them as random network fluctuations or congestion. This leads maintenance personnel to incorrectly adjust network bandwidth strategies, neglecting fundamental causes such as electromagnetic shielding failure or poor grounding, resulting in delayed fault warnings and false alarms. Summary of the Invention

[0004] This invention provides a real-time analysis and early warning system for vibration data of rotating pumps based on edge-side computing, which solves the technical problems mentioned in the background art.

[0005] This invention provides a real-time analysis and early warning system for vibration data of rotating pumps based on edge computing, comprising: a data monitoring module for capturing the error time sequence of the transmission link; and an edge processing module for constructing a dynamically rotating virtual electric vector space;

[0006] The edge processing module is configured to: dynamically adjust the rotation frequency of the virtual electric vector space using a gradient ascent algorithm, so as to maximize the clustering strength of the projected phase of the bit error time sequence mapped to the virtual electric vector space, thereby locking the optimal virtual frequency; and calculate the transmission bit error phase clustering discreteness characteristics based on the bit error probability density distribution under the optimal virtual frequency.

[0007] The edge processing module generates an anomaly warning signal for the transmission link based on the transmission error phase clustering discreteness characteristics based on the virtual electric vector space and a preset threshold.

[0008] The beneficial effects of this invention are as follows: by introducing a gradient ascent algorithm to drive the virtual electric vector space to automatically track and lock the instantaneous true frequency of the pump, and combining von Mises kernel density estimation and generalized Reynolds entropy analysis, the mathematical dimension of interference detection is elevated from simple low-dimensional statistics to high-dimensional manifold topology. Thus, even in complex environments with extremely low signal-to-noise ratios and real-time frequency drift, it can still accurately distinguish between random network congestion and electromagnetic coupling interference through significant changes in entropy values, greatly improving the robustness, specificity, and ability to capture weak fault characteristics of the edge-side early warning system. Attached Figure Description

[0009] Figure 1 This is a block diagram of the real-time analysis and early warning system for vibration data of rotating pumps based on edge-side computing according to the present invention. Detailed Implementation

[0010] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0011] like Figure 1 As shown, the real-time analysis and early warning system for vibration data of rotating pumps based on edge computing includes: a data monitoring module for capturing the error time sequence of the transmission link; and an edge processing module for constructing a dynamically rotating virtual electric vector space.

[0012] The edge processing module is configured to: dynamically adjust the rotation frequency of the virtual electric vector space using a gradient ascent algorithm, so as to maximize the clustering strength of the projected phase of the bit error time sequence mapped to the virtual electric vector space, thereby locking the optimal virtual frequency; and calculate the transmission bit error phase clustering discreteness characteristics based on the bit error probability density distribution under the optimal virtual frequency.

[0013] The edge processing module generates an anomaly warning signal for the transmission link based on the transmission error phase clustering discreteness characteristics based on the virtual electric vector space and a preset threshold.

[0014] It should be noted that capturing the error time sequence of the transmission link and constructing a differential error time vector is for subsequent virtual electric vector space mapping to eliminate the interference of the absolute time reference on the periodic feature analysis. Specifically, since the electromagnetic interference of pumps and motors usually acts on the communication link according to a specific periodic pattern, this pattern is reflected in the relative time interval of the error occurrence, rather than the absolute time of the error. However, the system clock of the edge-side equipment may drift, or there may be random offsets in the start time between different acquisition batches. If the absolute timestamp is directly used for spectrum analysis or tensor projection, the deviation of the zero point of time will be introduced into the subsequent calculation, resulting in the phase features being misaligned. Therefore, by constructing a differential error time vector, a high-pass filter is applied to the original time sequence, filtering out the DC component containing the start time deviation and retaining only the AC component representing the interference rhythm.

[0015] In a preferred embodiment, the low-level driver of the edge computing gateway collaborates with a field-programmable gate array (FPGA) or a network interface controller (NIC) to ensure microsecond-level timestamp accuracy. The system first initializes a sliding time window based on a circular buffer structure, which has a preset time span length. (For example, set to 1 second or 10 power frequency cycles) or preset capacity depth The data monitoring module monitors the status registers of the physical layer interface (e.g., RS-485 transceiver or Ethernet PHY chip) of the transmission link in real time via polling or interruption. Once the physical layer detects a Frame Check Sequence (FCS) error, a Cyclic Redundancy Check (CRC) error, or a physical layer retransmission request signal, the system immediately triggers a hardware interrupt, latching the current system high-precision counter value as an absolute timestamp. For example, within the current sliding time window, a sequence of raw error absolute timestamps arranged chronologically is captured. ,in This represents the total number of bit errors captured within the window. Indicates the first The absolute system time of each bit error, typically measured in microseconds. To construct a differential error timing vector that removes the effects of absolute time drift, the processor performs differential operations. Specifically, the differential interval is first defined. For the first in the sequence elements (of which) ),calculate For the first element of the sequence ( Considering the window boundary effect, it can be set to the difference between the last element of the previous window and the last element of the previous window, or set to 0 during cold start. The resulting difference data constitutes a one-dimensional vector. .vector Each element in This represents the silent time interval between two consecutive transmission failures caused by external electromagnetic interference when the transmission link is subjected to such interference. This set of time interval data preserves the topology of the operating frequency of the interference source (i.e., the pump). For example, if the pump operates at 50Hz and the interference occurs at a specific phase point, then... It will be presented in 20ms ( (Integer multiple features related to)

[0016] In some possible embodiments, to address the error aggregation phenomenon caused by extremely high frequency burst noise in industrial environments, a debounce preprocessing step can be added before constructing the differential error timing vector. Under certain extreme conditions, a single high-energy electromagnetic pulse may cause the physical layer interface to report multiple errors consecutively within a very short time (e.g., within a few microseconds), which should logically be considered as a single interference event. Therefore, a minimum resolution threshold is set. (For example, 100 microseconds). When the data monitoring module captures a new timestamp At that time, it will be compared with the timestamp of the previous record. Compare. Only when Only when the timestamp is valid will it be stored in the sliding time window; otherwise, it will be discarded as a redundant record of the same interfering event. After the de-jitter filtering is completed, the differential calculation in the preferred embodiment described above is performed using the filtered sequence.

[0017] In some possible embodiments, considering the computational resource limitations of edge devices, the sliding time window can employ an event-driven fixed-length FIFO (First-In-First-Out) queue mechanism instead of a fixed-length window. Specifically, the system allocates a fixed-length window in memory. (For example The system uses a queue of timestamps. Whenever a new physical layer event with a bit error is detected, the oldest timestamp is removed from the queue, the newest timestamp is added, and a local differential update calculation is immediately triggered. Specifically, the construction of the differential bit error timing vector is incremental; that is, each time only the difference between the newly added timestamp and the last timestamp is calculated, and one element of the vector is updated. This avoids periodic large-scale array operations, distributing the computational load evenly across each bit error occurrence, greatly reducing the requirements for instantaneous CPU performance, making it particularly suitable for low-power IoT edge gateway devices.

[0018] It should be noted that, in order to construct a virtual electric vector space capable of simulating the rotation mechanism of the internal magnetic field of the variable frequency drive, a mechanism of frequency domain coincidence is used to detect weak signals. Specifically, although the bit errors in the transmission link appear discrete and chaotic on the time axis, if these bit errors are caused by electromagnetic coupling due to the inverter's SVPWM (Space Vector Pulse Width Modulation) mechanism, then in a reference frame rotating at the inverter's operating frequency, the phase angles at which these bit errors occur should be highly concentrated or fixed. However, since the exact rotational speed of the pump is unknown, a high-dimensional space with variable rotational speed (i.e., virtual angular frequency) needs to be constructed. In the high-dimensional space, discrete time intervals are mapped to rotating vectors on a unit circle, and a target functional is defined as the metric. When the virtual frequency matches the real interference frequency, the vectors corresponding to all bit errors will point in approximately the same direction on the unit circle, at which point the magnitude of the vector sum (VectorSum) of these vectors is maximized, and the target functional reaches its extreme value; conversely, if the frequencies do not match, the vectors will randomly cancel each other out on the circumference, and the functional value approaches zero. This mechanism is similar to the principle of a stroboscope, which can filter out interference structures with specific periodicity from strong background noise.

[0019] In a preferred embodiment, the edge processing module first reads the differential error timing vector and restores it to a cumulative time series relative to zero points for positioning in a continuous rotation space. Subsequently, the system constructs a virtual electric vector space and performs functional calculations. The system sets a virtual angular frequency variable covering a preset power frequency range. For example, the scan range can be set to 0Hz to 100Hz. Simultaneously, the sector number characteristic coefficient of the space vector pulse width modulation mechanism can be set. For common three-phase six-arm inverters, this coefficient is typically set to 6, representing the six commutation sectors within one electrical cycle; for special twelve-pulse drivers, this coefficient can be set to 12. This applies to each virtual angular frequency to be tested. The system uses Euler's formula to map the bit error sequence onto the unit circle in the complex plane and calculates the target functional that measures the strength of phase clustering. The specific calculation formula for this objective functional is constructed as follows: In the formula, the parameters It represents the total number of bit error events captured within the sliding time window, and serves as a normalization function to eliminate the impact of fluctuations in the number of samples; The imaginary unit ( ); The aforementioned sector quantity characteristic coefficient (with a value of 6); The virtual angular frequency (unit: rad / s) of the current traversal. For the first The relative timestamps of each bit error event with respect to the start of the window (accumulated from the difference vector) (Obtained). The summation term in the formula. This means treating all bit errors as unit-length rotating vectors and performing vector superposition, with the outer modulus squared. This is used to quantize the total energy of the synthesized vector. This reflects the fact that at the interference frequency is Under the premise of [condition], the coherence of the error distribution. If A value close to 1 indicates that all errors occur at the same phase angle of the motor rotation, indicating a very strong correlation between interference.

[0020] In some possible embodiments, to accommodate low-cost FPGA or microcontroller edge gateways lacking floating-point units (FPUs), the process of constructing the virtual electric vector space can replace complex number operations using trigonometric orthogonal decomposition. Instead of directly calling complex exponential functions, the system decomposes the calculation of the target functional into two components: the real part (I-axis) and the imaginary part (Q-axis). Specifically, for each bit error time step... Calculate the cosine components Sine component Subsequently, for all the erroneous bits... and The total components are obtained by summing them up. and At this point, the objective functional Reconstructed into This implementation method can quickly complete trigonometric function calculations using a look-up table, reducing the algorithm's hardware computing power requirements.

[0021] In some possible embodiments, considering that different types of mechanical loads may be subject to non-electrical mechanical vibration disturbances (such as bearing failure), the sector number characteristic coefficient in this scheme... It can be configured as dynamically adjustable parameters or multimodal parallel computing parameters. The system not only calculates... The objective functional (corresponding to electrical interference) is also calculated in parallel. The fundamental frequency functional at time. When The corresponding functional value is significantly higher than In this case, the system can identify that the interference mainly originates from the fundamental frequency of the pump's mechanical rotation (such as vibration coupling caused by rotor imbalance), rather than the switching action of the frequency converter. This expands the dimension of the virtual electric vector space, enabling the early warning system not only to detect electromagnetic interference but also, to a certain extent, to differentiate between... The sensitivity of the signal is used to identify the attributes of the interference source, thereby enhancing the versatility of the technical solution and the depth of diagnosis.

[0022] It should be noted that after determining the initial frequency range of the virtual electric vector space, a gradient ascent algorithm is used to finely lock the virtual frequency, overcoming the frequency fence effect present in the traditional discrete scanning method. In actual operation, the pump speed is often non-integer (e.g., 49.87Hz) due to load fluctuations. If only a fixed step size scanning is used (e.g., 0.1Hz step size), the actual interference frequency may fall between two scanning points, causing the phase clustering features to diffuse due to frequency mismatch, and the calculated functional value cannot reach the true peak value. Therefore, the process of finding the optimal frequency is modeled as a convex optimization problem, and the target functional is regarded as a multidimensional surface. The derivative (gradient) is used to guide the search direction, so that the rotation frequency in the virtual space can automatically climb to the local maximum point along the steepest path of the functional surface, thereby achieving error-free phase locking of the pump's instantaneous speed.

[0023] In a preferred embodiment, the edge processing module performs a frequency iterative update process based on the analytic derivative of the complex plane. First, the system selects the target functional obtained in the preceding steps (such as coarse-grained full-spectrum scanning). The frequency corresponding to the local maximum point is used as the initial virtual angular frequency variable. Subsequently, the system enters an iterative loop, and at the [number]th [time / phase]... In this iteration, the objective functional with respect to the current virtual angular frequency is calculated. The partial derivatives. To avoid errors caused by complex numerical difference calculations, the precise value of the gradient is calculated directly using an analytical method. According to the chain rule, the objective functional... (in For weighted vector sum, (for its conjugate), its pair The formula for calculating partial derivatives is derived as follows: In the formula, This represents the current gradient value; This indicates taking the real part of a complex number; The imaginary unit; The sector quantity characteristic coefficient; and All of these are the time accumulation values ​​in the differential error timing vector; This represents the total number of samples. The first summation in the formula represents the conjugate composite vector at the current frequency, and the second summation represents the introduction of time weights. The weighted rotation vectors and their product reveal the sensitivity of phase cohesion to frequency variations. After obtaining the gradient, the system performs an update step: .in, The frequency of the next iteration; The preset learning rate determines the step size of frequency adjustments, and is typically a small positive number (e.g., ...). This ensures the iterative process remains stable and does not diverge. Specifically, if the current gradient has a positive sign, it means that increasing the frequency can improve the clustering strength, so the system increases the virtual rotation speed; otherwise, it decreases the rotation speed until the gradient approaches zero.

[0024] In some possible embodiments, to implement this functionality on computationally limited embedded edge gateways, numerical differential gradient methods can be used instead of analytical gradient methods. Instead of performing complex complex multiplications and differentiation operations, the processor uses a finite difference approximation to estimate the gradient direction. Specifically, in each iteration step, the system calculates the current frequency... and minute offset frequency (in The target functional value under extremely small perturbations (e.g., 0.01 Hz). and Subsequently, through the formula Calculate the approximate gradient. Although the mathematical accuracy of this method is slightly lower than that of the analytical method, it simplifies the complexity of code implementation, avoids dependence on complex derivative formulas, and can still effectively drive the virtual frequency to converge to the optimal operating point under the condition that the target functional is relatively smooth, fully meeting the real-time early warning requirements of the engineering site.

[0025] In some possible implementations, to address the complex electromagnetic environment that may exist in industrial settings, which could cause the target functional to exhibit multiple similar local extrema (i.e., a multimodal problem), this gradient ascent algorithm can introduce a momentum acceleration mechanism. The frequency update not only depends on the current gradient but also retains the update direction information from the previous step. Specifically, the current velocity variable is first calculated... ,in This is the momentum decay coefficient (e.g., 0.9). The previous update speed is used as the reference; then the update frequency is updated using that speed. This design principle of introducing a momentum term is similar to giving the virtual frequency search process inertia, enabling it to break through small local noise or flat regions by inertia, thereby avoiding getting trapped in false local optima and significantly accelerating the convergence speed from the initial frequency to the optimal locked frequency, thus improving the system's response sensitivity under dynamic disturbances.

[0026] It is important to note that during the gradient ascent iteration process, a termination mechanism must be introduced to determine the optimal virtual frequency to balance computational accuracy and system resource consumption. As the iteration points gradually approach the local maxima (i.e., the mountaintop) of the objective functional, the gradient (slope) of the objective function will gradually tend towards zero, causing the frequency update step size to naturally decrease. Without setting convergence conditions, the algorithm may get stuck in endless micro-jitter or invalid computation, wasting the valuable processor time of the edge gateway. Therefore, by monitoring the change in the frequency variable between two adjacent iterations, it is determined whether the system has determined the true value of the pump speed. When the frequency change is less than a very small threshold, the algorithm can be considered to have converged. The frequency at this point is the optimal virtual frequency that makes the bit error phase clusters most tightly, and this optimal virtual frequency reflects the electrical operating characteristics of the pump at the current moment.

[0027] In a preferred embodiment, the edge processing module performs a convergence check immediately after each frequency update operation. The system first reads the current... The virtual angular frequency variable calculated in the next iteration And the last time The virtual angular frequency variable in the next iteration Next, using the formula Calculate the absolute value of the difference between the two. This represents the correction magnitude made by the algorithm to the pump frequency estimation at this current step. Subsequently, the system will calculate... With the preset small convergence threshold Compare. The threshold. The setting is usually related to the system's physical frequency resolution. For example, if a frequency locking accuracy of 0.01Hz is required, then... Can be set to rad / s. Specifically, if If the iteration process converges, the system immediately sends a command to stop the gradient ascent algorithm loop and... Solidify into final output parameters (Optimal virtual frequency), which will be passed to the subsequent entropy calculation module. If If the algorithm has not yet reached the flat top of the functional surface, the system will continue with the next iteration. This precise control of when the iteration stops ensures that the system neither stops too early, leading to inaccurate frequency locking, nor causes computational delays due to excessive iteration.

[0028] In some possible implementations, to prevent the algorithm from failing to meet the difference convergence condition and falling into an infinite loop under extreme signal-to-noise ratio conditions (such as an overly flat target functional surface or the presence of high-frequency oscillating noise), a maximum number of iterations (MaxIterations) is introduced as a fallback mandatory termination condition. The system maintains a counter. .regardless Whether it is less than the threshold, as long as Exceeded the preset upper limit (For example, if set to 50 iterations), the system will forcibly stop the iteration and directly select the frequency corresponding to the maximum current objective functional value as the optimal virtual frequency. This enhances the robustness of the industrial field software.

[0029] In some possible implementations, the convergence criterion can be replaced by the rate of change of the independent variable with the gain rate of the dependent variable. Specifically, the system does not directly compare frequencies. Instead of comparing the change in the objective functional value before and after the iteration, we compare the objective functional value before and after the iteration. The extent of the improvement. System calculation. .like Less than the preset energy gain threshold This means that the improvement in phase clustering effect brought about by further adjusting the frequency is negligible and no longer has computational value, thus determining convergence and locking the frequency. Therefore, when the top of the target functional is very flat (i.e., the pump speed fluctuation is minimal and the phase alignment tolerance is high), the algorithm can avoid making meaningless fine adjustments near the extreme points, further improving the locking speed. This makes it suitable for ultra-low latency industrial control network environments with extremely high real-time requirements.

[0030] It should be noted that after successfully locking the optimal virtual frequency that maximizes phase clustering, a dimensionality boosting process is further performed from discrete time points to a continuous probability distribution to address the interference of discrete sampling sparsity on feature extraction. In actual transmission links, even with strong electromagnetic coupling, bit error events are random, manifesting as a finite number of discrete sampling points. Directly performing statistics on these sparse points (such as calculating variance) is highly susceptible to outliers and cannot accurately describe their distribution pattern (i.e., manifold structure). Therefore, by introducing kernel density estimation methods from manifold learning, particularly the von Mises kernel function suitable for periodic data, each discrete bit error moment is smoothed into a probability wave packet spreading in a circle. By superimposing all wave packets, the true distribution of bit errors in the phase space is reconstructed.

[0031] In a preferred embodiment, the edge processing module first utilizes the optimal virtual frequency. and preset sector number characteristic coefficients (Usually 6), each cumulative timestamp in the differential error timing vector Mapped onto the unit circle in virtual space. Specifically, calculate the... Phase angle corresponding to each bit error The calculation formula is: The modulo operation in this formula The linearly growing time axis was wound back to 0. On a closed circular toroidal manifold. Subsequently, the system selects the von Mises distribution function as the kernel function for all... Perform kernel density estimation to generate a continuous bit error phase probability density distribution function. The detailed calculation formula for this function is constructed as follows: In this formula, For the domain in A continuous phase variable within an interval represents any position on the circumference of a virtual space. This represents the total number of error samples within the sliding window. It is the base of the natural logarithm; The current position was measured. With the One actual error point angular distance between them; (Kappa) is the smoothing parameter (or concentration parameter), which is similar to the reciprocal of the variance in a Gaussian distribution. The larger the value, the sharper the peak of the generated distribution wave. The smaller the value, the smoother the distribution. An empirical value (such as 4.0) can be selected to balance resolution and smoothness; This is the Modified Bessel Function of Order Zero (MOBZ), which acts as a normalization constant to ensure that the sum of the probability integrals over the entire circle is 1. Using this formula, the system successfully constructs a continuous function describing the probability of bit errors as a function of the phase angle. The function exhibits a significant peak at the phase points of the interference set.

[0032] In some possible embodiments, considering the computation of Bessel functions in certain low-power IoT nodes... Since exponential functions can consume excessive CPU cycles, a high-resolution histogram smoothing fitting method can be used to approximate a rigorous kernel density estimate. The system will... arrive The circumferential interval is divided into equal parts. A tiny discrete sector (e.g.) (That is, each 1 degree is a sector). The system iterates through the phase angles of all bit errors. The histogram is then projected onto the corresponding sectors for counting, forming the original histogram. Subsequently, a circular convolution is performed on the original histogram data using a moving average filter of a preset width (e.g., a five-point moving average) or a Gaussian smoothing template. The resulting histogram envelope is then considered as an approximate probability density distribution function. .

[0033] In some possible embodiments, to improve the system's adaptive capability in non-steady-state noise environments, the smoothing parameter in the formula... It can be configured as a data-driven, dynamic variable rather than a fixed constant. The system computes the current sample set before performing KDE computation. Circular variance ,in The average composite vector length is given. Subsequently, the optimal value is dynamically calculated using a variant of the Rule of Thumb. Values, such as setting Specifically, when the error samples themselves are highly discrete (high noise), the noise level is automatically reduced. The value is increased to obtain a smoother estimate and avoid spurious noise peaks; when the error samples are highly concentrated (strong features), the value is automatically increased. This allows for sharper peak values, thereby improving the resolution of phase features. This enables the early warning system to maintain optimal feature extraction performance in industrial environments with drastically fluctuating signal-to-noise ratios.

[0034] It should be noted that after reconstructing the probability density distribution function of the bit error phase, the generalized entropy theory is further introduced to quantify the degree of disorder in this distribution, thereby utilizing the extreme sensitivity of information entropy to the distribution shape to distinguish between signals and noise. If the transmission link is subjected to modulation interference from the SVPWM mechanism, the phase distribution of the bit error... This will exhibit low entropy characteristics, meaning the probability mass is highly concentrated at a few specific phase points, leading to sharp peaks in the distribution function; conversely, if it is random network noise, the distribution will tend to be uniform, exhibiting high entropy characteristics. The chosen algorithm, based on the square integral of probability density, corresponds to the collision entropy in Rényi entropy. (Special case). Compared to the traditional Shannon entropy, this index is more sensitive to the peak structure in the probability distribution, and can more significantly widen the numerical difference between strong clusters (interference) and weak clusters (noise), thereby greatly improving the specificity of the early warning system and effectively suppressing false alarms.

[0035] In a preferred embodiment, the edge processing module performs an entropy feature calculation process based on numerical integration. Since the computer system processes discrete data, the system first processes continuous phase intervals... Discretize into Equally spaced sampling points (e.g.) ), step size System call probability density distribution function Calculate each discrete point function value at Subsequently, a squaring operation is performed to obtain the probability density squared sequence. Next, the definite integral values ​​are calculated using the Riemann Sum approximation method, that is, all... With step size The sum of the products yields the information potential energy or collision probability. Finally, the negative of the natural logarithm of this integral is calculated to obtain the generalized Rénix entropy. The complete mathematical expression of this calculation process is as follows: In the formula, This is the probability density function normalized by kernel density estimation; For integral resolution; squaring operation The weighting of high-probability regions has been strengthened; summation and multiplication Numerical integration has been completed; The operation maps the numerical range to a linear scale that aligns with human intuition (smaller entropy values ​​indicate greater order). If all bit errors completely overlap, It tends toward infinity (the maximum value in discrete cases). The error rate tends towards negative infinity; if the error rate is uniformly distributed, The integral value is , .

[0036] In some possible embodiments, to reduce the floating-point operation overhead of the embedded system, the direct information energy thresholding method can be used as an equivalent alternative to entropy calculation. The system omits the final logarithmic operation step and directly integrates the result. As an eigenvalue. Due to the logarithmic function It is a monotonically decreasing function, therefore, we can use it directly. Make a judgment. At this time, This represents the degree of order or clustering energy. Accordingly, the subsequent decision-making logic also needs to be adjusted to: if... If the value exceeds a certain high threshold, it is judged as an anomaly (strong interference clustering); if If the value is below the low threshold, it is considered normal (random noise).

[0037] In some possible embodiments, for applications requiring extremely high integration accuracy, the numerical integration step can use Simpson's rule instead of the simple Riemann sum. The system uses parabolic interpolation to approximate the curve shape of the squared probability density function. The calculation formula is adjusted as follows: This involves assigning different weighting coefficients (1, 4, 2) to the squared probability density values ​​of the first and last points, odd-indexed points, and even-indexed points, and then performing a weighted summation. This implementation method does not increase the number of sampling points. Under the premise of this, the truncation error of numerical integration can be significantly reduced, so that the calculated Reni entropy value can more accurately reflect the subtle structural changes in the phase distribution. It is particularly suitable for situations where the interference signal not only contains the fundamental frequency, but also contains a wealth of higher harmonics, resulting in an extremely complex probability distribution curve shape.

[0038] It should be noted that after calculating the generalized Reynolds entropy feature characterizing the dispersion of bit error phase clusters, the final anomaly detection stage begins. This establishes a dynamic reference frame based on environmental noise. Even with packet loss in a normal network transmission environment, the phase distribution should be random and disordered; therefore, the calculated entropy value should remain at a relatively high level (close to maximum entropy). Conversely, when the transmission link is subjected to periodic electromagnetic modulation by the inverter's SVPWM mechanism, the bit error phase is forcibly compressed to a few specific angles, causing a sharp drop in the system's entropy value. Therefore, the essence of this detection is to detect abnormal drops in entropy value. To adapt to the diverse electromagnetic noise levels in different industrial environments, this invention employs statistical thresholds based on historical data to ensure that the system automatically maintains optimal detection capabilities in both quiet laboratories and noisy factories.

[0039] In a preferred embodiment, the edge processing module executes a dynamic threshold generation and determination process based on sliding window statistics. The system maintains a historical normal baseline queue in memory to store a sequence of historical entropy values ​​that have been determined to be normal within a recent period (e.g., the past 30 minutes). Whenever a new decision needs to be made, the system first calculates the arithmetic mean of the data in the queue. and standard deviation Subsequently, based on the normal distribution... The Pauta Criterion constructs a dynamic decision threshold. The specific calculation formula is as follows: In the formula, This represents the average background entropy level under the current environment; Represents the fluctuation amplitude of background noise; This is the sensitivity coefficient, typically set to 3, meaning that only when the current entropy value deviates from the background mean by more than 3 standard deviations is it considered statistically significant. Since the core focus is on the drop in entropy value (i.e., an increase in orderliness), subtraction is used. After obtaining the threshold, the system uses the currently calculated transmission error phase clustering dispersion features based on the virtual electric vector space. and Compare them. If the inequality is satisfied... If the signal is not found to be congested, the system logic determines that the transmission link is not experiencing random congestion, but rather interference from a signal with strong clustering characteristics. This immediately generates a transmission link anomaly warning signal containing the interference frequency, intensity, and timestamp. Subsequently, to maintain the purity of the baseline queue, the signals deemed abnormal are... The value will not be updated into the historical baseline queue to prevent outliers from polluting the background statistical characteristics; conversely, if it is determined to be normal, the value will be updated into the queue, and the oldest data will be eliminated through the first-in-first-out mechanism to achieve adaptive drift of the threshold.

[0040] In some possible implementations, considering that in certain complex industrial scenarios, the entropy distribution of background noise may not follow a standard normal distribution (e.g., exhibiting a long-tailed or bimodal distribution), directly using the mean and standard deviation may lead to threshold estimation bias. Therefore, a non-parametric percentile thresholding method is employed. The system does not perform parameter estimation on historical queue data, but instead directly uses the values ​​in the queue... The historical entropy values ​​are sorted from smallest to largest. Then, the value at a specific percentile (e.g., the 1st percentile) is selected as the dynamic judgment threshold. .Right now .

[0041] In some possible embodiments, to prevent false alarms triggered by a sudden drop in entropy due to transient pulse interference, a temporal persistence check or hysteresis comparison mechanism can be added to the decision logic. The system has two states: monitoring state and alert state. Only when... continuous For example (Below the dynamic threshold) Only when the system switches from monitoring mode to alert mode and issues an alarm will the system switch to alert mode and issue an alarm. Alternatively, a dual threshold mechanism can be used: setting a lower trigger threshold. and a higher recovery threshold (in When the entropy value falls below An alarm is triggered only when the entropy value rises and exceeds a certain threshold. The alarm was only deactivated at that time.

[0042] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A rotating machine pump vibration data real-time analysis and early warning system based on edge side computing, characterized in that, The method comprises the following steps: a data monitoring module is configured to capture a sequence of error code time instants of a transmission link; and an edge processing module is configured to construct a dynamically rotating virtual electric vector space; the edge processing module is configured to dynamically adjust a rotation frequency of the virtual electric vector space by using a gradient ascent algorithm, so that a projection phase clustering intensity of the sequence of error code time instants mapped into the virtual electric vector space is maximized, thereby locking an optimal virtual frequency; and calculate a transmission error code phase clustering dispersion characteristic based on a virtual electric vector space based on an error code phase probability density distribution under the optimal virtual frequency; the edge processing module generates an abnormal early warning signal of the transmission link based on the transmission error code phase clustering dispersion characteristic based on the virtual electric vector space and a preset threshold value; wherein the rotating machine pump is a rotating machine pump based on space vector pulse width modulation control.

2. The edge-side computing based real-time analysis and early warning system for rotating machine pump vibration data according to claim 1, characterized in that, The method comprises the following steps: monitoring a frame check state of a transmission link physical layer interface of an edge computing gateway, and recording absolute time stamps when a physical layer check error or a physical layer retransmission event occurs; establishing a sliding time window, and extracting a plurality of absolute time stamps located in the sliding time window; calculating a differential time interval between adjacent two absolute time stamps, and combining all the differential time intervals in time sequence to construct a differential error code time sequence vector which is free from the influence of absolute time drift.

3. The edge-side computing based real-time analysis and early warning system for rotary machine pump vibration data according to claim 2, characterized in that, The method comprises the following steps: setting a virtual angular frequency variable covering a preset power frequency range; setting a sector number characteristic coefficient corresponding to a space vector pulse width modulation mechanism; for each virtual angular frequency variable, mapping the differential error code time sequence vector to a unit circle of a virtual electric vector space; and constructing a target functional which measures a phase clustering intensity, the target functional being used to represent a square value of a synthetic vector module length of the differential error code time sequence vector on the unit circle in combination with the sector number characteristic coefficient under a specific virtual angular frequency.

4. The edge-side computing-based real-time analysis and early warning system for rotary machine pump vibration data according to claim 3, characterized in that, The method comprises the following steps: selecting a frequency corresponding to a local maximum point of the target functional as an initial virtual angular frequency variable; calculating a partial derivative of the target functional with respect to the virtual angular frequency variable, the partial derivative representing a gradient direction of the target functional with respect to the virtual angular frequency variable; and performing an iterative updating step, using a product of a virtual angular frequency variable of a current iterative step and a preset learning rate and the partial derivative as a virtual angular frequency variable of a next iterative step, until a preset convergence condition is met.

5. The edge-side computing based real-time analysis and early warning system for rotary machine pump vibration data according to claim 4, characterized in that, The method comprises the following steps: calculating an absolute value of a difference value of the virtual angular frequency variables obtained by two adjacent iterative updating steps; judging whether the absolute value of the difference value is less than a preset small convergence threshold value; if the absolute value of the difference value is less than the small convergence threshold value, determining that the iterative process converges, and stopping the iterative updating of the gradient ascent algorithm; and taking the virtual angular frequency variable obtained by the final iterative updating as the optimal virtual frequency.

6. The edge-side computing based real-time analysis and early warning system for rotary machine pump vibration data according to claim 5, characterized in that, The error code phase probability density distribution at the optimal virtual frequency comprises: By using the optimal virtual frequency and the sector number characteristic coefficient, each time interval value in the differential error code time sequence vector is converted into a phase angle on a unit circle in a virtual electric vector space; a von Mises kernel function is selected as a kernel function for kernel density estimation, and kernel density estimation is performed on all the converted phase angles; and an error code phase probability density distribution function describing the distribution law of error code occurrence time in a rotating coordinate system driven by the optimal virtual frequency is constructed.

7. The edge-side computing based real-time analysis and early warning system for rotary machine pump vibration data according to claim 6, characterized in that, Calculating a transmission error code phase clustering dispersion characteristic based on a virtual electric vector space comprises: A square operation is performed on the error code phase probability density distribution function to obtain a probability density square function; a definite integral value of the probability density square function in an entire circumferential phase interval is calculated; a reciprocal of a natural logarithm of the definite integral value is calculated to obtain a generalized Renyi entropy value; and the generalized Renyi entropy value is taken as the transmission error code phase clustering dispersion characteristic based on the virtual electric vector space.

8. The edge-side computing based real-time analysis and early warning system for rotary machine pump vibration data according to claim 7, characterized in that, Based on the transmission error code phase clustering dispersion characteristic based on the virtual electric vector space and a preset threshold value, an abnormal early warning signal of a transmission link is generated, comprising: A mean value and a standard deviation of the generalized Renyi entropy value in a historical normal operation time period are counted to calculate a dynamic determination threshold value based on an environmental noise base; the transmission error code phase clustering dispersion characteristic based on the virtual electric vector space calculated at present is compared with the dynamic determination threshold value in a numerical value; if the transmission error code phase clustering dispersion characteristic based on the virtual electric vector space is smaller than the dynamic determination threshold value, it is determined that the transmission link is modulated by periodic electromagnetic interference of a variable frequency driving device, it is confirmed that the transmission link is abnormal, and the abnormal early warning signal of the transmission link is output.

Images