A tuning fork position limiting signal abnormality recognition method and system

Abstract

This invention discloses a method and system for identifying abnormal tuning fork limit signals, relating to the field of data processing for industrial sensor signals. The method includes: applying an active excitation cutoff operation to the piezoelectric drive chip of the tuning fork sensor at a preset trigger frequency and cutting off the electrical excitation signal input within the excitation cutoff time window; acquiring the free decay waveform of the piezoelectric receiving chip; extracting the energy decay asymmetry, the absolute total energy integral value across the window, and the phase drift variation coefficient to construct a multi-dimensional feature vector; inputting this vector into a machine learning-based media state classification model for prediction, obtaining classification results for empty state, normal full state, and vortex street self-excited full abnormal state; outputting the corresponding limit signal based on the classification results and triggering an alarm when an abnormality occurs. This invention solves the problem of missed full-material alarms caused by Karman vortex street resonance lock-up that may be caused by high-velocity media scouring the tuning fork.

Description

Technical Field

[0001] This invention relates to the field of data processing of industrial sensor signals, and more specifically, to a method and system for identifying abnormal tuning fork limit signals. Background Technology

[0002] Tuning fork sensors, indispensable level measurement instruments in modern industrial process control, are widely used in storage containers and transmission pipelines in industries such as chemical, petroleum, and pharmaceutical. Their underlying physical architecture typically comprises symmetrically distributed mechanical tuning fork bodies, with embedded piezoelectric drive and receiver chips. Under normal operating conditions, the electronic control circuit inputs an alternating electrical excitation signal of a specific resonant frequency to the piezoelectric drive chip, driving the tuning fork body to maintain a high-amplitude mechanical resonance in the air. When the liquid medium in the pipeline rises and completely submerges the tuning fork body, due to the fluid's much higher viscous damping characteristics compared to air, a large amount of the mechanical kinetic energy of the tuning fork body's vibration is absorbed, causing a momentary collapse in the amplitude of the voltage signal induced at the receiver. Traditional level instruments rely on this sudden change in amplitude or frequency to determine the state of the medium.

[0003] As industrial control evolves towards higher precision and more complex operating conditions, simple hardware threshold comparisons are gradually shifting towards digital and intelligent pattern recognition. For example, Chinese Patent Publication No. CN117312769A discloses a BiLSTM-based method for detecting anomalies in IoT time-series data. This method acquires and preprocesses time-series data from IoT sensing devices, inputs it into a pre-trained deep learning model to output predicted time-series data, and then uses the comparison between the residual and a threshold to intelligently label the anomaly data. Another example is Chinese Patent Publication No. CN115309736B, which discloses a time-series data anomaly detection method based on a self-supervised learning multi-head attention network. This method targets time-series data collected from multiple industrial sensors, using data augmentation pre-training and a multi-head attention network structure for deep feature mapping, thereby achieving high-precision anomaly isolation in industrial time-series data. These schemes utilizing advanced data processing and algorithmic models for time-series signal analysis provide important references for data processing in modern precision instruments.

[0004] However, in actual high-flow-rate industrial pipelines, existing technologies face significant failure risks. When high-velocity media laterally scour a non-streamlined tuning fork, it creates a Karman vortex street by alternating shedding behind the fork. When the flow velocity reaches a certain critical range, the alternating frequency of vortex shedding approaches the tuning fork's inherent mechanical resonant frequency, triggering a severe frequency lock-in phenomenon. Once lock-in occurs, external fluid vortices will generate continuous forced excitation forces near the tuning fork's inherent frequency, forcing the tuning fork to maintain high-amplitude, violent vibrations similar to those in air, even when completely submerged. This can cause the system to lock into an empty state, leading to extremely dangerous overflow and leakage accidents. Furthermore, fluid conditions are not always stable; drastic fluctuations in flow velocity cause the formation and disappearance of Karman vortex streets to occur in a highly unstable critical transition zone. This unsteady fluctuation causes the instrument's underlying judgment logic to frequently jump between normal submersion and abnormal resonance, resulting in high-frequency oscillations in the terminal alarm output. Moreover, if the system is constantly subjected to complex high-frequency algorithm calculations to prevent such occasional anomalies, it will significantly increase the instrument system's power consumption and shorten hardware lifespan. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and system for identifying abnormal tuning fork limit signals, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for identifying abnormal tuning fork limit signals includes the following steps: An active excitation cutoff operation is applied to the piezoelectric drive chip of the tuning fork sensor at a preset trigger frequency. The active excitation cutoff operation includes cutting off the electrical excitation signal input of the piezoelectric drive chip within a preset excitation cutoff time window; and acquiring the free attenuation waveform of the piezoelectric receiving chip of the tuning fork sensor within the preset excitation cutoff time window. Based on the above, a multidimensional feature vector is constructed; the multidimensional feature vector is input into a pre-trained machine learning-based medium state classification model for classification prediction to obtain the current medium state classification result, which includes the empty state classification result, the normal full state classification result, and the vortex sheave self-excited full abnormal state classification result. Based on the current medium state classification result, a corresponding limit signal is output. When the current medium state classification result is the vortex street self-excited full abnormal state classification result, it is determined that the current tuning fork signal has a full material leakage alarm abnormality, and a corresponding abnormal alarm limit signal is output.

[0007] Preferably, the process of constructing the multidimensional feature vector includes: The free decay waveform is divided into a first half and a second half. The energy integral value of the first half waveform is compared with the energy integral value of the second half waveform to generate the energy decay asymmetry. The free decay waveform is integrated in the entire time domain over the entire preset excitation time window to generate the absolute total energy integral value of the entire window. The timestamps of each zero-crossing point in the free decay waveform are obtained, the standard deviation and average value of the time interval between adjacent zero-crossing points are calculated, and the ratio of the standard deviation and the average value is calculated to generate the phase drift variation coefficient. The energy decay asymmetry, the absolute total energy integral value of the full window, and the phase drift variation coefficient are concatenated into vectors to generate the multidimensional feature vector.

[0008] Preferably, the machine learning-based media state classification model is a fully connected neural network model.

[0009] Preferably, the training process of the fully connected neural network model includes: Obtain multiple sample free decay waveforms under historical excitation-off states; extract the sample energy decay asymmetry, the sample full-window absolute total energy integral value, and the sample phase drift variation coefficient from each sample free decay waveform and combine them into a sample feature vector; Each of the sample feature vectors is labeled with a corresponding empty state label, a normal full state label, and a vortex shedding full abnormal state label. The initial fully connected neural network model is iteratively trained using the sample feature vector labeled with the tags until the preset convergence condition is met, resulting in a fully connected neural network model that has been trained.

[0010] Preferably, the method further includes a daily limit output step: when the current medium state classification result is the empty state classification result, the limit signal is kept as an incomplete state signal; when the current medium state classification result is the normal full state classification result, the corresponding full normal state signal is output.

[0011] Preferably, the identification method includes a fixed-period detection mode and a symptom-triggered detection mode; When in the fixed-cycle detection mode, the preset trigger frequency is a constant basic detection frequency; When in the symptom trigger detection mode, before the step of applying active excitation interruption operation to the piezoelectric drive chip of the tuning fork sensor at a preset trigger frequency, a transient disturbance capture process based on dual-source frequency traction symptom triggering is also combined.

[0012] Preferably, the transient disturbance capture process based on dual-source frequency traction symptom triggering specifically includes the following steps: Under normal electrically excited driving conditions, the relative phase slip ratio between the electric excitation signal input to the piezoelectric driving chip and the voltage signal acquired from the piezoelectric receiving chip is continuously extracted; The voltage signal is envelope demodulated, and the amplitude envelope modulation depth of the voltage signal in the low-frequency band is calculated. The relative phase slip ratio of the drive and the modulation depth of the amplitude envelope are combined into an excitation feature vector, which is input into a preset machine learning-based dual-source excitation coherence evaluation model for binary classification prediction. The corresponding excitation identification result is output, which includes fluid turbulence excitation result and normal stationary result. When the disturbance identification result is the normal and stable result, the preset trigger frequency is set to zero so that the system is exempt from performing the active excitation cut-off operation; When the disturbance identification result is the fluid turbulence disturbance result, it is taken as a forced resonance precursor of the Karman vortex street self-excited full abnormal state. The current output state of the limit signal remains unchanged, and the preset trigger frequency is increased to the preset encrypted trigger frequency to trigger the application of the active excitation cut-off operation.

[0013] Preferably, the method further includes a transition state self-excited oscillation elimination step, specifically including: Within a series of preset excitation time windows, the probability distribution vectors corresponding to each classification state generated by the machine learning-based medium state classification model when outputting the classification results are obtained, and a time-series probability matrix is ​​constructed. The time series probability matrix is ​​input into a preset state transition probability correction model for Markov time series smoothing, and the corrected state decision value is output. The corrected state determination decision value is used to replace the current medium state classification result of the single prediction output in order to execute the subsequent output of the limit signal.

[0014] Preferably, the training process of the state transition probability correction model includes: Under the condition of fluctuating flow velocity, obtain multiple sets of historical time series probability matrices that transition from the normal full state classification result to the vortex sheave self-excited full abnormal state classification result, and use them as transition state samples. Label each of the transition state samples with the corresponding macroscopic steady-state target label; The initial state transition probability correction model is fitted with parameters using the transition state samples labeled with the macroscopic steady-state target to establish state transition damping weights for the self-excited locking and unlocking process, thus obtaining the trained state transition probability correction model.

[0015] This invention also discloses a system for implementing the above-mentioned method for identifying abnormal tuning fork limit signals, comprising: An active excitation interruption module is used to apply an active excitation interruption operation to the piezoelectric drive chip of the tuning fork sensor at a preset trigger frequency. The active excitation interruption operation includes cutting off the electrical excitation signal input of the piezoelectric drive chip within a preset excitation interruption time window. The signal acquisition module is used to acquire the free attenuation waveform of the piezoelectric receiving crystal of the tuning fork sensor within the preset excitation time window; The feature extraction module is used to construct a multidimensional feature vector based on the free decay waveform; The state identification module is used to input the multidimensional feature vector into a pre-trained machine learning-based medium state classification model for classification prediction to obtain the current medium state classification result, wherein the current medium state classification result includes the empty state classification result, the normal full state classification result, and the vortex shedding full abnormal state classification result. The signal output module is used to output a corresponding limit signal according to the current medium state classification result, and when the current medium state classification result is the vortex street self-excited full abnormal state classification result, it determines that the current tuning fork signal has a full material leakage alarm abnormality and outputs a corresponding abnormal alarm limit signal.

[0016] The advantage of this invention over the prior art is that it forces the tuning fork into a free decay state without electrical signal interference by performing an active excitation cut-off operation at a preset trigger frequency, and extracts three core features from it: energy decay asymmetry, the absolute total energy integral value of the whole window, and the phase drift variation coefficient, thus realizing multi-dimensional state decoupling from a physical essence. The energy decay asymmetry utilizes the ratio of the energy integral of the first half to the second half of the free decay waveform. Its physical mechanism lies in the fact that when vortices in the Karman vortex street detach, the fluid acts like an external mechanical pump, continuously applying periodic thrust to the tuning fork. This continuous energy replenishment effect causes the second half of the waveform to deviate from the natural exponential decay law, thus inducing ratio distortion. The absolute total energy integral value across the entire window utilizes the high viscosity and damping characteristics of the fluid. Under normal submerged conditions, the liquid drastically suppresses the tuning fork amplitude, causing the integral value across the entire window to approach zero, thus effectively distinguishing between low-energy submersion and high-energy resonance. The phase drift variation coefficient captures the microscopic phase noise caused by unsteady turbulence. When the fluid engulfs the tuning fork, the randomness of the vortex forces the tuning fork to deviate from the rigid beat determined by its inherent stiffness, producing fluctuating zero-crossing point jitter. The organic combination of these three features accurately identifies vortex street lock-up anomalies and eliminates underreporting of full-load conditions caused by the Karman vortex street.

[0017] Building upon this foundation, the present invention introduces a transient excitation capture process triggered by dual-source frequency traction symptoms, solving the early warning problem of abnormal operating conditions from the perspective of dynamic coupling. In the critical region of the Karman vortex street lock-in, the tuning fork is subjected to the dual effects of an electronic drive signal and external eddy current forcing. According to the frequency traction effect in nonlinear vibration, when the two frequencies approach each other, the system will generate significant beat frequency oscillations, reflected in the voltage signal as an increase in the amplitude envelope modulation depth and a periodic shift in the relative phase of the drive and receiver. By utilizing a dual-source excitation coherence evaluation model, the present invention can capture this beat frequency symptom caused by frequency traction before the lock-in state is fully formed. It then achieves high-precision encrypted verification by dynamically increasing the trigger frequency, while simultaneously stopping power-consuming excitation under stable operating conditions. This represents a leap from blind periodic detection to intelligent symptom defense, resolving the contradiction between power consumption and reliability.

[0018] Furthermore, to address the jitter in critical state determination caused by flow velocity fluctuations, this invention further integrates a transitional self-excited oscillation elimination mechanism. Because the Karman vortex street exhibits nonlinear bifurcation characteristics at the critical Reynolds number, the system is prone to switching between locked and unlocked states. Instead of directly employing a single-prediction hard threshold determination, this invention introduces Markov time-series smoothing processing. By constructing a time-series probability matrix, the state transition lag from physics is introduced into the algorithm logic. This process is equivalent to introducing a layer of virtual state damping into the logical determination, utilizing the inertial transition characteristics of the system state in the time dimension to filter out pseudo-abrupt signals caused by instantaneous flow velocity fluctuations. This smoothing processing effectively smooths out the determination oscillations in the critical region, making the output limit signal stable and highly confident, thereby thoroughly ensuring the operational safety of downstream actuators under complex fluctuating conditions and avoiding contact arcing and equipment wear caused by frequent switching actions. Attached Figure Description

[0019] Figure 1 This is a structural block diagram of the overall system of the present invention, which sequentially shows the flow of the five core hardware or logic modules that implement the method: active tripping module, signal acquisition module, feature extraction module, state identification module, and the final signal output module.

[0020] Figure 2 This is a waveform diagram of the active excitation cutoff operation and free decay of the present invention, showing the dynamic process of the waveform changing from constant amplitude high frequency to damped free decay after the electrical excitation signal is cut off within the set excitation cutoff time window.

[0021] Figure 3 This is a schematic diagram of the principle of energy attenuation asymmetry feature extraction of the present invention. The attenuation waveform is divided into two parts in the picture, and the energy integral values ​​of the first half and the second half are compared with the bar chart below to obtain the ratio.

[0022] Figure 4This invention provides a phase drift variation coefficient extraction diagram, depicting an enlarged waveform that clearly identifies the zero-crossing points on the waveform and the time intervals (Δt1~Δt4) between adjacent zero-crossing points, used to calculate the ratio of the standard deviation to the mean.

[0023] Figure 5 This is a diagram of the medium state classification model based on a fully connected neural network, which shows the hierarchical relationship of three different working conditions (empty, full, and abnormal) after three-dimensional features are input into the hidden layer network of the model.

[0024] Figure 6 This is a diagram of the dual-source frequency traction symptom triggering process of the present invention, which shows the logical judgment path of the system extracting the envelope and phase shift from the two signals, performing coherence evaluation, and triggering encrypted detection when fluid turbulence symptoms are detected.

[0025] Figure 7 This is a physical image of the abnormal state of the Karman vortex street caused by fluid turbulence, which simulates the physical phenomenon of the Karman vortex street formed by the alternating shedding of the fork teeth when the fluid washes over them from a top-down perspective, revealing the cause of forced resonance.

[0026] Figure 8 This is a schematic diagram of the principle of eliminating transitional self-excited oscillations in this invention. It shows how to use a continuous time probability matrix and damping weights to smoothly correct the judgment value of the jump spike to a stable decision through a Markov model. Detailed Implementation

[0027] The embodiments of the present invention will be described below with reference to the accompanying drawings. The hardware circuits, sampling circuits, processors, and communication interfaces in each embodiment can be implemented using devices commonly found in industrial level instruments, such as microcontrollers, DSPs, FPGAs, analog switches, operational amplifiers, ADC sampling chips, relay output circuits, transistor output circuits, or bus communication circuits. Any device capable of performing active tripping, waveform acquisition, feature calculation, state recognition, and limit signal output can serve as a vehicle for implementing the present invention.

[0028] The overall system of this invention may include an active tripping module, a signal acquisition module, a feature extraction module, a state identification module, and a signal output module, such as... Figure 1 As shown. The active excitation cut-off module is connected to the piezoelectric drive chip of the tuning fork sensor to control the conduction and cut-off of the electrical excitation signal; the signal acquisition module is connected to the piezoelectric receiver chip to acquire the free decay waveform; the feature extraction module is used to perform feature calculation on the free decay waveform; the state identification module outputs the medium state classification result based on the machine learning model; the signal output module generates an incomplete state signal, a full material normal state signal, or an abnormal alarm limit signal according to the classification result.

[0029] During normal testing, the electronic control circuit first inputs an alternating electrical excitation signal to the piezoelectric drive chip that matches the mechanical resonant frequency of the tuning fork, causing the tuning fork body to be in a stable vibration state. Subsequently, the active de-energizing module performs an active de-energizing operation according to a preset trigger frequency. Active de-energizing operation refers to temporarily cutting off the electrical excitation signal input to the piezoelectric drive chip within a preset de-energizing time window, preventing the tuning fork from receiving continuous energy from the electronic drive source. The preset de-energizing time window can be determined based on the tuning fork's natural frequency, the degree of dielectric damping, and the sampling processing capability; it is typically between 5ms and 500ms, preferably between 20ms and 200ms. The preset trigger frequency can be between 0.01Hz and 10Hz, and in a fixed-period testing mode, it can be between 0.02Hz and 2Hz, for example, testing once every 10s or once every 2s. In another embodiment, the preset trigger frequency can also be dynamically adjusted according to the degree of eddy-induced risk. Specifically, when the system determines, based on historical operating data, current flow velocity range, medium viscosity, pipeline pressure, tuning fork installation posture, and dual-source frequency traction symptoms, that it is currently in a low-risk time window where vortex street self-excitation is unlikely to occur, the preset trigger frequency can be reduced to a low-frequency inspection frequency, such as 0.001Hz to 0.05Hz, or only long-interval health confirmation-type interruption detection can be retained to reduce disturbance to the normal limit detection process and reduce system power consumption. When subsequent monitoring detects an increase in the relative phase slip ratio of drive and receiver, an increase in amplitude envelope modulation depth, a flow velocity entering the vortex street locking sensitive range, or an increase in the probability of anomalies under similar historical operating conditions, the preset trigger frequency is restored to the basic detection frequency or increased to the encrypted trigger frequency. In this way, the active interruption operation does not need to be executed at a fixed high frequency all the time, but can reduce the number of detections when the risk of vortex-induced excitation is low, and quickly increase the detection density when vortex-induced excitation symptoms appear, thereby balancing identification reliability, low power consumption, and hardware lifespan.

[0030] After active excitation cutoff, the tuning fork enters a free decay state. At this point, the signal acquired by the piezoelectric receiver chip is no longer the steady-state oscillation forcibly maintained by the drive circuit, but rather a decaying waveform formed by the tuning fork mechanical system releasing energy in the current medium environment, such as... Figure 2 As shown. The purpose of this design is to remove the masking effect of the electrical excitation signal on the received signal, so that the three physical mechanisms of liquid damping, low air damping, and external energy replenishment from the Karman vortex street are fully exposed in the waveform. Under normal full-load conditions, the liquid has strong viscous damping, and the free decay collapses quickly; under empty conditions, the air damping is small, and the free decay is relatively smooth; under the abnormal full-load condition of the vortex street self-excited vortex, even if the electrical excitation is cut off, the fluid vortex may still continuously apply periodic mechanical force to the tuning fork, so the decay process will exhibit abnormal tailing, phase jitter, or local energy replenishment phenomena.

[0031] The signal acquisition module acquires the voltage signal output by the piezoelectric receiver chip within a preset excitation time window and forms a free decay waveform. Before acquisition, high-frequency noise can be suppressed by analog low-pass filtering or digital filtering, or DC bias removal can be performed in the digital domain to ensure that subsequent characteristics primarily reflect the tuning fork vibration itself. The sampling frequency can be set according to the principle of covering the tuning fork resonant frequency and its decay details, for example, not less than 20 times the main vibration frequency of the tuning fork.

[0032] The feature extraction module extracts the energy attenuation asymmetry, the total absolute energy integral value of the full window, and the phase drift variation coefficient from the free attenuation waveform, and concatenates the three features into a multi-dimensional feature vector.

[0033] When extracting the energy decay asymmetry, the free decay waveform within the preset excitation-off time window is divided into a first half and a second half in chronological order. The energy integral values ​​of the first half and the second half are calculated separately. Then, the ratio of the first half and the second half energy integral values ​​is calculated. Figure 3 As shown. The energy integral value can be obtained by summing the squared amplitudes or by summing the rectified absolute amplitudes as the equivalent energy representation, but the calculation method should be consistent within the same instrument or model version. Under normal air conditions, the first and second halves exhibit a relatively stable natural decay relationship; under normal full-load conditions, the energy in both the first and second halves decreases significantly; under abnormal vortex street self-excited full-load conditions, the second half may be abnormally large due to continuous energy replenishment by fluid vortices, causing the energy decay asymmetry to deviate from the normal range. To avoid division by zero when the energy in the second half approaches 0, a minimum resolvable energy can be set as a protection value, and the larger of the second half energy integral value and this protection value can be used in the ratio calculation. This protection value can be taken as 0.001 to 0.01 times the average energy of the entire window of no-load stable oscillation.

[0034] When extracting the total absolute energy integral value for the entire window, the free decay waveform is subjected to full-time domain absolute integration over the entire preset excitation-off time window. In practice, the absolute amplitude after removing the DC bias at each sampling point can be accumulated, or the absolute amplitude can be multiplied by the sampling interval before accumulation. If the system uniformly uses the square of the amplitude as the energy representation during the calibration phase, the square amplitude can also be accumulated and converted to the equivalent total energy value for the entire window using the same calibration coefficient, but this should be consistent with the training phase during the online identification phase. This feature is mainly used to distinguish between low-energy full-load and high-energy abnormal resonance. Under normal full-load conditions, the strong damping of the liquid will rapidly dissipate the mechanical energy of the tuning fork, and the total absolute energy integral value for the entire window will be significantly reduced; under the abnormal full-load condition of vortex street self-excited, although the tuning fork is submerged in liquid, the forced vibration of the external vortex street will keep the receiver at a relatively high energy level, so the integral value will not approach the low-energy range as quickly as under normal full-load conditions.

[0035] When extracting the phase drift variation coefficient, first obtain the timestamps of each zero-crossing point in the free decay waveform, then calculate the time interval between adjacent zero-crossing points, and subsequently calculate the standard deviation and average of these time intervals. Then, perform a ratio operation between the standard deviation and the average. In a specific embodiment, this can be the standard deviation divided by the average. Figure 4 As shown. The zero-crossing timestamp can be determined by the sign change of adjacent sampling points, and the timing accuracy can be improved by combining linear interpolation. The phase drift variation coefficient reflects the beat stability during the free decay process. In the empty state, the tuning fork mainly decays according to its own natural frequency, and the zero-crossing interval is relatively uniform; in the normal full-load state, the amplitude decays quickly, and the number of zero-crossings may decrease; in the vortex street self-excited full-load abnormal state, the random disturbance of the fluid vortex will cause the zero-crossings to drift erratically, resulting in an increase in the variation coefficient. If there are fewer than 3 effective zero-crossings within the preset excitation time window, the phase drift variation coefficient can be set to the median value of the phase drift variation coefficient corresponding to the normal full-load state in the training samples or the preset default value, while the strong damping low-energy characteristics of the window are still reflected by the absolute total energy integral value of the entire window. This processing can avoid the instability of calculation due to insufficient number of zero-crossings, and will not change the feature dimension of the input model.

[0036] After obtaining the three features, the feature extraction module concatenates the energy attenuation asymmetry, the total absolute energy integral value across the entire window, and the phase drift variation coefficient into a 3D feature vector in a fixed order. To improve the model's adaptability to different batches of tuning forks, different installation methods, and different media conditions, the mean and scale coefficient of each feature can be recorded during the model training phase and normalized according to the same rules during the online recognition phase. Normalization does not change the physical meaning of the features; it only prevents the total absolute energy integral value across the entire window from unreasonably dominating the model's judgment due to its large dimension.

[0037] The state identification module inputs multi-dimensional feature vectors into a pre-trained machine learning-based medium state classification model to obtain the current medium state classification result. The medium state classification results include empty state classification results, normal full state classification results, and vortex shedding self-excited full abnormal state classification results. The machine learning-based medium state classification model can employ a fully connected neural network model, such as... Figure 5 As shown in the figure, the input layer of this model includes three input nodes, corresponding to the energy decay asymmetry, the absolute total energy integral value of the full window, and the phase drift variation coefficient, respectively; the hidden layer can be set to 1 to 3 layers, each containing 8 to 64 neurons, and the activation function can be ReLU, tanh, or other common nonlinear functions; the output layer includes three output nodes, corresponding to the empty state, the normal full state, and the vortex shedding full anomalous state, and the probability distribution of the three states is formed by probabilistic output.

[0038] The training process for the medium state classification model can be carried out as follows: First, acquire multiple sample free decay waveforms under historical excitation-discontinuation states. The samples should cover empty states, normal full states, and vortex street self-excited full abnormal states. Empty state samples can be collected when the tuning fork is not in contact with the medium; normal full state samples can be collected under static or low-flow-rate full-material conditions; vortex street self-excited full abnormal state samples can be collected in high-flow-rate pipelines when the fluid laterally scours the tuning fork teeth and generates a Karman vortex street, such as... Figure 7 As shown. Subsequently, the sample energy decay asymmetry, the sample full-window absolute total energy integral value, and the sample phase drift variation coefficient are extracted from the free decay waveform of each sample and combined into a sample feature vector. Then, the sample feature vector is labeled with empty state label, normal full state label, or vortex street self-excited full abnormal state label.

[0039] After labeling, the initial fully connected neural network model is iteratively trained using the labeled sample feature vectors. Training can employ the cross-entropy loss function and optimization algorithms such as Adam and SGD. The learning rate can be between 0.0001 and 0.01, the batch size between 16 and 256, and the number of training epochs between 50 and 1000. Preset convergence conditions may include training loss falling below a preset loss threshold, validation set accuracy reaching 95% or higher, or validation set loss no longer decreasing after 10 to 50 consecutive epochs. After training, the model parameters are embedded into the instrument controller, edge computing module, or host computer for online classification and prediction.

[0040] The signal output module outputs corresponding limit signals based on the current medium state classification result. When the classification result is an empty state, the limit signal remains in a partially submerged state, indicating that the tuning fork is not effectively submerged by the medium. When the classification result is a normal full state, a full-material normal state signal is output, indicating that the tuning fork has been submerged by the medium and there has been no vortex street self-excitation lockup. When the classification result is a vortex street self-excitation full abnormal state, it is determined that the current tuning fork signal has a full-material underreporting abnormality, and an abnormal alarm limit signal is output. The abnormal alarm limit signal can be linked with the full-material safety protection logic, causing the downstream actuator to handle the situation as a full-material danger state, such as closing the feed valve, stopping pumping, or triggering an audible and visual alarm. In this way, even if the traditional amplitude determination logic mistakenly identifies an empty material due to high amplitude, this invention can still identify the real risk based on the free decay abnormal characteristics.

[0041] In a further embodiment, the present invention may include a fixed-period detection mode and a symptom-triggered detection mode. In the fixed-period detection mode, the preset trigger frequency is a constant basic detection frequency. The basic detection frequency can be between 0.02Hz and 2Hz, suitable for scenarios where operating conditions are relatively stable and power consumption is not particularly sensitive. This mode is simple to implement; the controller only needs to trigger an active interruption operation according to the timer period to periodically obtain the free decay waveform and perform state identification.

[0042] In symptom-triggered detection mode, before performing active interruption operation, a transient disturbance capture process based on dual-source frequency-driven symptom triggering is first performed, such as... Figure 6 As shown. Here, the dual source refers to the electrical excitation signal applied to the piezoelectric drive chip by the electronic control circuit, and the external mechanical excitation applied to the tuning fork body by the fluid vortex street. When the frequency of fluid vortex shedding gradually approaches the natural frequency of the tuning fork, the tuning fork will be affected by both electronic drive and external vortex forcing force. Phase slip and envelope fluctuations will appear in the voltage signal at the receiving end. These phenomena can be regarded as signs of the formation of Karman vortex street lock-in.

[0043] The transient disturbance capture process can be implemented as follows: Under normal electrically excited driving conditions, the controller continuously acquires the electrical excitation signal input to the piezoelectric drive chip and the voltage signal output from the piezoelectric receiver chip. Using existing techniques such as lock-in amplification, digital phase-sensitive detection, PLL phase tracking, or Hilbert transform, the relative phase between the drive and receiver signals is extracted, and the phase slip ratio over time is calculated. The more pronounced the relative phase slip ratio, the more likely the tuning fork vibration beat is subject to external fluid forcing.

[0044] Simultaneously, the voltage signal acquired by the piezoelectric receiving chip is envelope demodulated, and the amplitude envelope modulation depth in the low-frequency band is calculated. Envelope demodulation can be achieved using methods such as rectified low-pass, Hilbert transform, or digital peak tracking. The amplitude envelope modulation depth is used to characterize the strength of beat frequency oscillations. When the vortex shelving frequency is not yet fully locked but is already close to the tuning fork's natural frequency, a tension will occur between the external eddy current forcing force and the electronic drive, causing the voltage signal envelope to periodically bulge and fall, thus increasing the amplitude envelope modulation depth.

[0045] Subsequently, the relative phase slip rate of the driving and receiving forces and the amplitude envelope modulation depth are combined into an excitation feature vector, which is then input into a pre-defined machine learning-based dual-source excitation coherence assessment model. This model performs binary classification prediction, outputting either the fluid turbulence excitation result or the normal stationary result. The dual-source excitation coherence assessment model can employ logistic regression, support vector machine, random forest, or a small fully connected neural network. If a small fully connected neural network is used, its input layer can include two input nodes, the hidden layer can include one to two layers, each with four to sixteen neurons, and the output layer includes two nodes, corresponding to fluid turbulence excitation and normal stationary flow, respectively. Model training samples can be derived from the phase slip rate and envelope modulation depth under historical conventional driving conditions, and labels can be generated by combining real flow velocities, manual annotations, or subsequent excitation interruption identification results.

[0046] In the symptom-triggered detection mode, when the disturbance identification result is normal and stable, the preset trigger frequency for the active excitation cutoff operation is set to 0, thus preventing the system from performing the active excitation cutoff operation. Setting it to 0 here means disabling or suspending the active excitation cutoff timing trigger; it does not mean the drive signal stops, nor does it affect the tuning fork continuing to be in the normal electrically excited detection state. This reduces unnecessary excitation cutoff operations and model calculations, lowers power consumption, and extends hardware lifespan.

[0047] When the disturbance identification result is a fluid turbulence disturbance, it is considered a precursor to forced resonance in the abnormal state of a Karman vortex street self-excited full state. At this time, the system does not immediately change the limit signal, maintaining its current output state to avoid false alarms based solely on the symptom. Simultaneously, the preset trigger frequency is increased to a preset encrypted trigger frequency to apply active excitation interruption operations more frequently and acquire free decay waveforms. The encrypted trigger frequency can range from 1Hz to 50Hz, preferably from 2Hz to 20Hz. In this way, the system performs high-frequency fine-tuning only when the vortex street risk increases, reducing intervention in stable states, thus achieving a balance between reliability and low power consumption.

[0048] In another embodiment, the invention further includes a transitional self-excited oscillation elimination step. This step addresses the problem of frequent jumps in classification results between normal full state and abnormal vortex street self-excited full state caused by flow velocity fluctuations. Because the Karman vortex street exhibits nonlinear bifurcation characteristics near the critical flow velocity, the tuning fork may experience a process of locking, unlocking, and re-locking within a short period. If the single-shot model prediction result is directly used as the limit output, relays, valves, or alarms may operate at high frequencies, causing contact arcing, equipment wear, or erroneous interlocking of the control system.

[0049] In the transition state self-excited oscillation elimination step, the system acquires the probability distribution vectors corresponding to each classification state generated when the medium state classification model outputs classification results within multiple consecutive preset excitation interruption time windows, and constructs a time-series probability matrix. The number of consecutive windows can be 3 to 20, preferably 5 to 10. Each row of the time-series probability matrix corresponds to one excitation interruption detection, and each column corresponds to the probability of an empty state, a normal full state, or an abnormal vortex shedding self-excited full state. This matrix retains the process information of state changes over time, rather than only retaining the result with the highest probability each time.

[0050] Subsequently, the temporal probability matrix is ​​input into a preset state transition probability correction model for Markov temporal smoothing, such as... Figure 8As shown. The state transition probability correction model can include a state transition matrix, state holding damping weights, and observation confidence weights. The state transition matrix represents the probability tendency of transitions between empty states, normal full states, and vortex street self-excited full anomalous states; the state holding damping weights reflect that the physical state will not change instantaneously without inertia; and the observation confidence weights measure the impact of the current model output probability on the final decision. In actual calculations, common Markov processing methods such as forward filtering, Viterbi path search, or recursive Bayesian smoothing can be used. After processing, the model outputs the corrected state decision value. The corrected state decision value can be represented as a corrected category result or as a corrected state probability vector. The system uses the corrected state decision value to replace the current medium state classification result of a single prediction output before executing subsequent limit signal outputs.

[0051] The training of the state transition probability correction model can be performed as follows: First, obtain multiple sets of historical time-series probability matrices showing the transition from the normal full-state classification results to the vortex street self-excited full-abnormal state classification results under flow velocity fluctuation conditions, and use these matrices as transition state samples. The transition state samples should cover conditions such as increased flow velocity, decreased flow velocity, short-term vortex street disturbance, continuous lockup, and short-term unlocking. Then, label each transition state sample with a corresponding macroscopic steady-state target label. The macroscopic steady-state target label is determined based on duration, manual confirmation, flow record, or safety control requirements, and can be labeled as empty state, normal full state, or vortex street self-excited full-abnormal state. For the critical transition segment between the normal full state and the vortex street self-excited full-abnormal state, the macroscopic steady-state target label should prioritize reflecting the stable output result that should be adopted for safety control during this period, rather than simply following a single instantaneous prediction result.

[0052] Subsequently, the parameters of the initial state transition probability correction model were fitted using transition state samples labeled with macroscopic steady-state targets. The fitting objective was to establish state transition damping weights for the self-excited lock-up and unlocking processes, making the model more inclined to maintain its original stable state when faced with short-term spikes, and able to switch to an anomaly alarm state in a timely manner when faced with sustained high-probability anomalies. The state transition damping weights can be set from 0.5 to 0.99; higher values ​​indicate that the system is less likely to change its output due to short-term probability jumps. The observation confidence weights can be set from 0.1 to 1; higher values ​​indicate that the current classification model output has a greater impact on the final decision. After training, the state transition probability correction model was embedded into the controller and participated in online judgment together with the medium state classification model.

[0053] In terms of system implementation, the active excitation cutoff module can be implemented using analog switches, MOSFET switches, or drive enable control terminals. During excitation cutoff, the controller shuts down the excitation channel of the piezoelectric driver chip, and if necessary, places the drive terminal in a high-impedance state to avoid the residual impedance of the drive circuit having an additional impact on the free decay of the tuning fork. The signal acquisition module can include a charge amplifier, voltage amplifier, anti-aliasing filter, and ADC sampling unit to convert the weak voltage signal generated by the piezoelectric receiving chip into a digital waveform. The feature extraction module can be implemented by an MCU, DSP, or FPGA to complete waveform segmentation, energy integration, absolute integration, zero-crossing detection, and feature vector concatenation. The state identification module can deploy a fully connected neural network model and can use floating-point inference, fixed-point inference, or lookup table approximation inference depending on the device capabilities. The signal output module can output switch quantities, relay contacts, transistor signals, 4mA to 20mA current signals, or communication signals such as RS485, CAN, and industrial Ethernet.

[0054] Through the above implementation methods, when a tuning fork is subjected to high-velocity liquid scouring and generates a Karman vortex street, this invention no longer relies solely on the conventional amplitude magnitude to determine the material level. Instead, it actively cuts off the electrical excitation and observes the free decay behavior of the tuning fork under conditions without electrical drive interference. Energy decay asymmetry is used to identify external eddy current energy replenishment, the full-window absolute total energy integral value is used to distinguish between strong liquid damping and abnormally high-energy resonance, and the phase drift variation coefficient is used to capture beat instability caused by turbulence. Combined with symptom triggering and Markov timing smoothing, this invention can improve the ability to identify full-load under-reporting anomalies while reducing power consumption and suppressing output oscillations at critical flow rates.

[0055] In another simplified embodiment, instead of explicitly extracting the energy attenuation asymmetry, the total absolute energy integral value across the entire window, and the phase drift variation coefficient from the free decay waveform, the detection sequence acquired within a preset excitation-off time window can be directly input into a machine learning-based dielectric state classification model. This detection sequence can be the original voltage sampling sequence output by the piezoelectric receiver chip, or a standardized sequence after DC bias removal, amplitude normalization, bandpass filtering, or resampling. Since the sampling length and sampling frequency of different tuning fork sensors may differ, the detection sequence can be uniformly truncated or padded to a fixed length, such as 128 to 4096 points, to meet the dimensionality requirements of the model's input layer. In this case, the dielectric state classification model can employ a one-dimensional convolutional neural network, a recurrent neural network, a Transformer coding network, or an end-to-end classification model formed by a combination of one-dimensional convolutional layers and fully connected layers. This model automatically learns implicit features in the free decay waveform, such as decay rate, energy tail, zero-crossing jitter, and local energy replenishment, through convolutional kernels, temporal memory units, or attention weights within the network. It then outputs classification results for empty states, normal full states, and vortex sheave self-excited full abnormal states. The advantage of this approach is that it does not rely on manually defined feature extraction rules and can preserve subtle transient information in the free decay waveform. It is suitable for applications with a sufficient number of samples, high controller computing power, or where further improvements in the accuracy of complex condition recognition are required.

[0056] The training process of this end-to-end model can include: acquiring multiple sample detection sequences under historical excitation states; performing uniform length processing, filtering, and normalization on each sample detection sequence; labeling each sample detection sequence with empty state, normal full state, and vortex shedding full abnormal state labels; inputting the labeled sample detection sequences into the initial end-to-end classification model for iterative training, so that the state classification results output by the model gradually approach the corresponding labels; stopping training and obtaining the trained end-to-end medium state classification model when the training loss is lower than a preset threshold, the validation set classification accuracy reaches a preset requirement, or the validation set loss no longer decreases after multiple consecutive rounds. After training, in the online detection stage, new detection sequences are input into the model according to the same preprocessing rules to directly obtain the current medium state classification result, and limit signal output is executed accordingly.

[0057] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying abnormal tuning fork limit signals, characterized in that, Includes the following steps: An active excitation cutoff operation is applied to the piezoelectric drive chip of the tuning fork sensor at a preset trigger frequency. The active excitation cutoff operation includes cutting off the electrical excitation signal input of the piezoelectric drive chip within a preset excitation cutoff time window; and acquiring the free attenuation waveform of the piezoelectric receiving chip of the tuning fork sensor within the preset excitation cutoff time window. Construct a multidimensional feature vector based on the free decay waveform; The multidimensional feature vector is input into a pre-trained machine learning-based medium state classification model for classification prediction to obtain the current medium state classification result. The current medium state classification result includes the empty state classification result, the normal full state classification result, and the vortex sheave self-excited full abnormal state classification result. Based on the current medium state classification result, a corresponding limit signal is output. When the current medium state classification result is the vortex street self-excited full abnormal state classification result, it is determined that the current tuning fork signal has a full material leakage alarm abnormality, and a corresponding abnormal alarm limit signal is output.

2. The method for identifying abnormal tuning fork limit signals according to claim 1, characterized in that, The specific steps for constructing a multidimensional feature vector based on the free decay waveform include: The free decay waveform is divided into a first half and a second half. The energy integral value of the first half waveform is compared with the energy integral value of the second half waveform to generate the energy decay asymmetry. The free decay waveform is integrated in the entire time domain over the entire preset excitation time window to generate the absolute total energy integral value of the entire window. The timestamps of each zero-crossing point in the free decay waveform are obtained, the standard deviation and average value of the time interval between adjacent zero-crossing points are calculated, and the ratio of the standard deviation and the average value is calculated to generate the phase drift variation coefficient. The energy decay asymmetry, the absolute total energy integral value of the full window, and the phase drift variation coefficient are concatenated into vectors to generate the multidimensional feature vector.

3. The method for identifying abnormal tuning fork limit signals according to claim 2, characterized in that, The machine learning-based media state classification model is a fully connected neural network model.

4. The method for identifying abnormal tuning fork limit signals according to claim 3, characterized in that, The training process of the fully connected neural network model includes: Obtain multiple sample free decay waveforms under historical excitation-off states; extract the sample energy decay asymmetry, the sample full-window absolute total energy integral value, and the sample phase drift variation coefficient from each sample free decay waveform and combine them into a sample feature vector; Each of the sample feature vectors is labeled with a corresponding empty state label, a normal full state label, and a vortex shedding full abnormal state label. The initial fully connected neural network model is iteratively trained using the sample feature vector labeled with the tags until the preset convergence condition is met, resulting in a fully connected neural network model that has been trained.

5. The method for identifying abnormal tuning fork limit signals according to claim 1, characterized in that, It also includes a daily limit output step: when the current medium status classification result is the empty status classification result, the limit signal is kept as an incomplete status signal; when the current medium status classification result is the normal full status classification result, the corresponding full normal status signal is output.

6. The method for identifying abnormal tuning fork limit signals according to claim 1, characterized in that, The identification method includes a fixed-period detection mode and a symptom-triggered detection mode; When in the fixed-cycle detection mode, the preset trigger frequency is a constant basic detection frequency; When in the symptom trigger detection mode, before the step of applying active excitation interruption operation to the piezoelectric drive chip of the tuning fork sensor at a preset trigger frequency, a transient disturbance capture process based on dual-source frequency traction symptom triggering is also combined.

7. The method for identifying abnormal tuning fork limit signals according to claim 6, characterized in that, The transient disturbance acquisition process based on dual-source frequency traction symptom triggering specifically includes the following steps: Under normal electrically excited driving conditions, the relative phase slip ratio between the electric excitation signal input to the piezoelectric driving chip and the voltage signal acquired from the piezoelectric receiving chip is continuously extracted; The voltage signal is envelope demodulated, and the amplitude envelope modulation depth of the voltage signal in the low-frequency band is calculated. The relative phase slip ratio of the drive and the modulation depth of the amplitude envelope are combined into an excitation feature vector, which is input into a preset machine learning-based dual-source excitation coherence evaluation model for binary classification prediction. The corresponding excitation identification result is output, which includes fluid turbulence excitation result and normal stationary result. When the disturbance identification result is the normal and stable result, the preset trigger frequency is set to zero so that the system is exempt from performing the active excitation cut-off operation; When the disturbance identification result is the fluid turbulence disturbance result, it is taken as a forced resonance precursor of the Karman vortex street self-excited full abnormal state. The current output state of the limit signal remains unchanged, and the preset trigger frequency is increased to the preset encrypted trigger frequency to trigger the application of the active excitation cut-off operation.

8. The method for identifying abnormal tuning fork limit signals according to claim 1, characterized in that, It also includes a transition state self-excited oscillation elimination step, specifically including: Within a series of preset excitation time windows, the probability distribution vectors corresponding to each classification state generated by the machine learning-based medium state classification model when outputting the classification results are obtained, and a time-series probability matrix is ​​constructed. The time series probability matrix is ​​input into a preset state transition probability correction model for Markov time series smoothing, and the corrected state decision value is output. The corrected state determination decision value is used to replace the current medium state classification result of the single prediction output in order to execute the subsequent output of the limit signal.

9. The method for identifying abnormal tuning fork limit signals according to claim 8, characterized in that, The training process of the state transition probability correction model includes: Under the condition of fluctuating flow velocity, obtain multiple sets of historical time series probability matrices that transition from the normal full state classification result to the vortex sheave self-excited full abnormal state classification result, and use them as transition state samples. Label each of the transition state samples with the corresponding macroscopic steady-state target label; The initial state transition probability correction model is fitted with parameters using the transition state samples labeled with the macroscopic steady-state target to establish state transition damping weights for the self-excited locking and unlocking process, thus obtaining the trained state transition probability correction model.

10. A system for implementing the method for identifying abnormal tuning fork limit signals according to any one of claims 1 to 9, characterized in that, include: An active excitation interruption module is used to apply an active excitation interruption operation to the piezoelectric drive chip of the tuning fork sensor at a preset trigger frequency. The active excitation interruption operation includes cutting off the electrical excitation signal input of the piezoelectric drive chip within a preset excitation interruption time window. The signal acquisition module is used to acquire the free attenuation waveform of the piezoelectric receiving crystal of the tuning fork sensor within the preset excitation time window; The feature extraction module is used to construct a multidimensional feature vector based on the free decay waveform; The state identification module is used to input the multidimensional feature vector into a pre-trained machine learning-based medium state classification model for classification prediction to obtain the current medium state classification result, wherein the current medium state classification result includes the empty state classification result, the normal full state classification result, and the vortex shedding full abnormal state classification result. The signal output module is used to output a corresponding limit signal according to the current medium state classification result, and when the current medium state classification result is the vortex street self-excited full abnormal state classification result, it determines that the current tuning fork signal has a full material leakage alarm abnormality and outputs a corresponding abnormal alarm limit signal.

Images