A Dynamic Weighing Verification Method for Food Flour Based on Pressure Waveform Analysis
By acquiring the micro-shear stress and micro-amplitude resonance signals of the weighing platform, and combining the knowledge graph of powder flowability and the graph attention network model, the problem of error identification and verification of powder dynamic weighing technology under unsteady flow conditions was solved, and high-precision real-time error verification and adaptive verification were achieved.
Patent Information
- Application Number
- CN202611280230.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-22
- Publication Date
- 2026-11-17
AI Technical Summary
Existing dynamic powder weighing technology lacks accuracy in error identification and verification under unsteady flow conditions, making it difficult to adapt to complex flow behavior. Traditional methods also lack spatiotemporal adaptability and physical interpretability.
By acquiring the micro-shear stress response signal and micro-amplitude resonance signal of the weighing platform, and combining the powder flowability knowledge graph for physical property parameter matching, a flow state classification label set is generated. Then, the waveform segmentation engine and graph attention network model are used for error discrimination to achieve real-time error verification.
It enables rapid capture and accurate attribution of non-steady-state flow disturbances, has good online applicability and adaptability to various powder materials, and reduces engineering deployment costs.
Smart Images

Figure CN123111440A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of powder dynamic weighing error identification and pressure waveform analysis technology, and in particular to a dynamic weighing and metering verification method for food powder based on pressure waveform analysis. Background Technology
[0002] With the popularization of industrial automation and intelligent manufacturing, dynamic powder weighing technology has become a key link in continuous feeding and process monitoring in the food, pharmaceutical, and chemical industries. Most existing mainstream weighing systems use pressure sensors or strain gauge bridges as core detection elements, acquiring and analyzing pressure waveforms during powder flow to achieve quality determination and error correction. In terms of current technology, the industry mainly follows two paths: one is waveform analysis based on fixed time windows, suitable for static or low-speed conditions; the other is extracting waveform features through frequency domain transformation and multi-sensor fusion to enhance the system's adaptability to complex flows. In recent years, intelligent sensor arrays, multimodal data fusion, and neural network error recognition algorithms have become increasingly popular. Some high-end systems have introduced technologies such as digital twins and deep image recognition to improve measurement accuracy and robustness in dynamic scenarios.
[0003] However, existing technologies face significant challenges under unsteady flow conditions. When powders exhibit complex behaviors such as agglomeration and slippage, dispersion and retention, airbag rupture, bridging collapse, or high-speed gushing, the accuracy of error identification and verification drops sharply, making it difficult to meet the requirements of high-precision dynamic metering. Key shortcomings include: traditional feature extraction methods rely on fixed windows and static thresholds, failing to dynamically adapt to the nonlinear time-varying characteristics of the flow state, leading to unstable capture of key error causes and a high risk of missed or misjudged cases; while multi-sensor fusion can improve signal dimensionality, it relies too heavily on parameter tuning and physical interpretation, resulting in high integration complexity and insufficient real-time performance; and while deep learning models possess some generalization ability, they lack physical mechanism-driven approaches, have poor interpretability, and struggle to support online closed-loop compensation.
[0004] Therefore, the field of dynamic powder weighing urgently needs an error identification and verification technology that breaks through the limitations of fixed windows and possesses both spatiotemporal adaptability and physical interpretability. This technology can perceive the mechanical behavior of the powder interface in real time, perform semantic segmentation based on flow state transition points using a flowability knowledge graph, dynamically allocate feature extraction strategies, and utilize graph neural networks to fuse segmented features with causal relationships, ultimately achieving accurate, traceable, and condition-adaptive dynamic error verification. This invention addresses the need for waveform semantic segmentation and error feature anchoring driven by powder flow state, constructing a full-link, high real-time, and highly interpretable dynamic weighing error identification system, providing technical support for industrial intelligent metering and process safety monitoring. Summary of the Invention
[0005] This application provides a dynamic weighing and calibration method for food powder based on pressure waveform analysis, which aims to solve the problem of reading deviation caused by complex flow instability mechanisms such as gas-solid coupling oscillation and intermittent bridging when weighing powder under unsteady conditions.
[0006] The dynamic weighing and measurement verification method for food flour based on pressure waveform analysis provided in this application specifically includes:
[0007] The micro-shear stress response signal and micro-amplitude resonance signal of the load-bearing area of the weighing platform are obtained and combined into a raw mechanical dataset;
[0008] Based on a pre-set knowledge graph of powder flowability, the physical property parameters of the original mechanical dataset are matched to generate a set of flow state classification labels for the dominant mode of the current working condition.
[0009] Based on the flow state classification label set, the waveform segmentation engine is invoked to semantically segment the original mechanical dataset using the physical inflection points in the pressure waveform as boundaries, generating a sequence of pressure waveform segmentation units with clear physical meaning.
[0010] For each segment in the pressure waveform segmentation unit sequence, feature extraction is performed, a multi-dimensional state attribute vector including the pressure decay rate curvature extremum or the initial unloading slope is calculated, and an attributed directed graph structure carrying the temporal and mechanical coupling relationship is constructed.
[0011] The attributed directed graph structure is input into a pre-trained state-waveform-error map attention network model, and an error-sensitive feature weighted aggregation operation is performed to generate an error discrimination vector with state traceability.
[0012] Based on the error discrimination vector, the dynamic weighing error caused by unsteady flow is analyzed, the real-time error correction compensation factor is calculated, and the adaptive verification result for correcting the weighing reading is output.
[0013] The dynamic weighing and measurement verification method for food flour based on pressure waveform analysis provided in this application has the following beneficial effects:
[0014] By constructing a three-layer causal identification mechanism based on flow state, waveform morphology, and error causes, pressure waveforms are treated as a direct mechanical response carrier of powder dynamic behavior. A shear stress sensor array and micro-vibration probe are combined to achieve synchronous sensing of multi-source physical fields. Furthermore, relying on a powder flowability knowledge graph to match the dominant flow pattern under the current working condition, prior guidance is provided for differentiated waveform analysis. This effectively overcomes the shortcomings of traditional methods, such as delayed response, high misjudgment rate, and lack of physical explanation support under sudden flow disturbances. Simultaneously, this application introduces a flow state-driven waveform segmentation engine and a graph-structured feature aggregation mechanism. Based on the specific morphology of pressure waveforms under different flow states, segmentation logic and exclusive feature sets with clear mechanical meaning are defined. All segmentation units are organized into attributed directed graph structures, and adaptive weighted fusion is performed through a graph attention network. This results in error discrimination vectors with both high sensitivity and traceability, capable of backtracking to the specific flow event type that caused the error. This solves the technical bottleneck of black-box models being unable to attribute causes. Moreover, this architecture only requires changing the powder entries in the knowledge graph to adapt to new materials, reducing engineering deployment costs.
[0015] The above technical solutions jointly construct a dynamic weighing error identification system based on physical mechanisms and a closed-loop linkage of perception, identification, and attribution. Without relying on digital twin simulation or complex multi-sensor fusion modeling, it achieves rapid capture and accurate attribution of unsteady flow disturbances, and has good online applicability and wide adaptability to various powder materials. Attached Figure Description
[0016] Figure 1 This is the main flowchart of the dynamic weighing and measurement verification method for food powder based on pressure waveform analysis.
[0017] Figure 2 This is a sub-flowchart of a dynamic weighing and measurement verification method for food powder based on pressure waveform analysis.
[0018] Figure 3 This is another sub-flowchart of the dynamic weighing and measurement verification method for food powder based on pressure waveform analysis. Detailed Implementation
[0019] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0020] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.
[0021] like Figure 1 As shown, this application provides a dynamic weighing and measurement verification method for food flour based on pressure waveform analysis, specifically including:
[0022] S1: Obtain the micro-shear stress response signal and micro-amplitude resonance signal of the load-bearing area of the weighing platform, and simultaneously fuse the shear stress response signal and micro-amplitude resonance signal into a mechanical raw dataset.
[0023] S2: Based on the pre-set knowledge graph of powder flowability, the physical property parameters of the original mechanical dataset are matched to generate a set of flow state classification labels that represent the dominant mode of the current working condition.
[0024] S3: Based on the flow state classification label set, the waveform segmentation engine is called to semantically segment the original mechanical dataset using the physical inflection points in the pressure waveform as boundaries, generating a sequence of pressure waveform segmentation units with clear physical meaning.
[0025] S4: For each segment in the pressure waveform segmentation unit sequence, perform a feature extraction strategy, calculate a multi-dimensional state attribute vector including the pressure decay rate curvature extremum or the initial unloading slope, and construct an attributed directed graph structure carrying the temporal and mechanical coupling relationship.
[0026] S5: Input the attributed directed graph structure into the pre-trained state-waveform-error map attention network model, perform error-sensitive feature weighted aggregation operation, and generate an error discrimination vector with state traceability.
[0027] S6: Based on the error discrimination vector, analyze the causes of dynamic weighing errors caused by unsteady flow, calculate the real-time error correction compensation factor, and output the adaptive verification result for correcting the weighing reading.
[0028] like Figure 2 As shown, step S1 involves acquiring the micro-shear stress response signal and micro-amplitude resonance signal of the load-bearing area of the weighing platform, and simultaneously fusing the shear stress response signal and micro-amplitude resonance signal into a raw mechanical dataset. Specifically, this includes:
[0029] S1.1: High-frequency sampling control is performed on the micro shear stress sensor array in the key load-bearing area of the symmetrical weighing platform to obtain discrete shear stress response time series data characterizing the slip and agglomeration behavior of powder particles.
[0030] A micro-MEMS shear stress sensor array is deployed in the key load-bearing area of the weighing platform. The array is distributed in a grid pattern along the material flow direction to cover the main stress surface.
[0031] A constant current source is applied to the sensor array and a low-noise preamplifier circuit is configured to convert the weak resistance change signal into an analog voltage quantity.
[0032] A high-precision analog-to-digital converter is used to perform high-frequency synchronous sampling of the voltage analog quantity, with the sampling frequency set to 5kHz to capture millisecond-level particle slip transients.
[0033] A sliding window mean filter is applied to the discrete sampled data, with the window length set to 10 sampling points to suppress high-frequency electronic noise interference.
[0034] The filtered data is fused using the sensor spatial coordinate matrix to eliminate single-point measurement bias and generate a global shear stress response sequence.
[0035] Through the above processing method, the original electrical signal is transformed into discrete shear stress response time series data that characterizes the slip and agglomeration behavior of powder particles, thereby achieving high signal-to-noise ratio mechanical feature extraction.
[0036] For example, a MEMS shear stress sensor with a range of 0-5N and a sensitivity of 2mV / V was selected and placed at five key points at the four corners and center of the bottom of the weighing platform. The ADC sampling rate was set to 5000Hz and the resolution to 16-bit. A 10-point moving average filter was applied to the acquired signal to remove electronic noise higher than 500Hz. Assuming that the output voltage of the center sensor is 10.2mV at a certain moment, the shear stress value is calculated to be 0.51N after calibration. Combining the data from the surrounding sensors, the average shear stress of the current cross-section is calculated to be 0.48N using a weighted average algorithm, forming a stress time series with a timestamp accurate to 0.2ms, which significantly improves the ability to capture small particle slip events.
[0037] S1.2: Perform transient signal capture operation on the high frame rate micro-vibration probes arranged at the nodes of the platform structure to extract discrete micro-amplitude resonance signal time series data reflecting the oscillation characteristics of powder impact and gas-solid coupling.
[0038] At the critical mechanical transmission nodes of the weighing platform structure, a high-sensitivity piezoelectric micro-vibration probe array is deployed. The probe has wide-band response characteristics to cover the low-frequency resonance caused by powder impact and the high-frequency flutter caused by gas-solid coupling.
[0039] The analog voltage signal output by the micro-vibration probe is pre-charge amplified to convert the high-impedance charge signal into a low-impedance voltage signal, and an anti-aliasing filter is applied to suppress noise interference above the Nyquist frequency.
[0040] The filtered voltage signal is synchronously sampled using a high-speed analog-to-digital converter with a sampling frequency of at least 20kHz to ensure that the waveform details of millisecond-level transient impact events can be captured, generating a discretized original vibration amplitude time series.
[0041] Baseline drift correction was performed on the original vibration amplitude time series, and the DC component offset caused by changes in ambient temperature or mechanical creep was removed by sliding window mean subtraction to obtain a pure vibration signal with zero mean.
[0042] The pure vibration signal is decomposed into time and frequency based on the Hilbert-Huang transform algorithm to extract the instantaneous frequency and instantaneous amplitude envelope, and to identify the pulse peak representing the impact of powder particles on the platform and the continuous oscillation mode representing the airflow disturbance.
[0043] The energy integral value of the instantaneous amplitude envelope is calculated as a quantitative indicator reflecting the impact intensity of the powder. At the same time, the entropy value of the instantaneous frequency distribution is statistically analyzed as a characteristic parameter characterizing the complexity of gas-solid coupling oscillation.
[0044] The extracted pulse peaks, continuous oscillation modes, energy integral values, and frequency distribution entropy values are recombined in time stamp order to form discrete micro-amplitude resonance signal time series data that reflects the oscillation characteristics of powder impact and gas-solid coupling.
[0045] By employing the aforementioned high-frequency sampling, baseline correction, and time-frequency feature extraction processing methods, the physical vibration response from the previous step is transformed into discrete micro-amplitude resonance signal time series data containing transient impact and steady-state oscillation information. This enables high-fidelity capture of subtle mechanical disturbances in unsteady flow, providing high-precision vibration dimension input for subsequent multi-modal signal synchronous fusion.
[0046] S1.3: Based on a high-precision global clock source, nanosecond-level timestamp alignment processing is performed on the discrete shear stress response time series data and the discrete micro-amplitude resonance signal time series data to generate a dual-channel synchronous mechanical signal pair with a strict time sequence correspondence.
[0047] The discrete shear stress response time series data obtained from S1.1 and the discrete micro-amplitude resonance signal time series data obtained from S1.2 are received as the original input signals of the multimode to be aligned.
[0048] The high-precision global clock source module is invoked to read the current system reference timestamp. This clock source uses a temperature-controlled crystal oscillator (OCXO) as the frequency reference to ensure long-term frequency stability is better than the 1e-9 order of magnitude.
[0049] A nanosecond-level absolute time stamp is attached to each sampling point in the shear stress response signal stream, and the signal arrival time is captured by a hardware-triggered interrupt mechanism to eliminate random jitter errors introduced by software scheduling delays.
[0050] Perform the same hardware-level timestamp operation on the micro-amplitude resonant signal stream to ensure that the two signals have a unified time reference system at the source of acquisition.
[0051] A time resampling algorithm based on linear interpolation is constructed to calculate the time deviation between the sampling time of the two signals and the global clock reference. This deviation is defined as the difference between the actual sampling time and the ideal synchronization time.
[0052] Based on the calculated time offset, fractional delay filtering is performed on signal channels with high sampling rates or phase lag, and time shift compensation with subsampling point precision is achieved using an FIR filter.
[0053] The time-shift compensated shear stress signal and the micro-amplitude resonance signal are strictly matched according to the timestamp to generate a one-to-one corresponding dual-channel synchronous mechanical signal pair.
[0054] By using nanosecond-level timestamp alignment and fractional delay filtering, the discrete signals independently acquired in the previous step are transformed into dual-channel synchronous mechanical signal pairs with strict temporal correspondence, realizing the accurate fusion of multimodal data at the microscopic time scale and providing a phase-distortion-free data foundation for subsequent spatial registration.
[0055] For example, the global clock source frequency is set to 100MHz, with a period resolution of 10ns. The shear stress sensor has a sampling rate of 50kHz, and the micro-vibration probe has a sampling rate of 100kHz. Within a certain acquisition window, the Nth point of the shear stress signal is detected at 1000.000020s, and the corresponding time for the micro-vibration signal is 1000.000015s, resulting in a calculated time offset of 5μs. To address this offset, a 64-order FIR fractional delay filter is designed, with the group delay set to the proportion of sampling points corresponding to 5μs. After filtering the micro-vibration signal, its phase lag is accurately corrected. In the final generated dual-channel signal pair, the shear stress value and the micro-vibration value at any time t correspond to the state of the same physical instant, with the time synchronization error controlled within ±2ns. This significantly eliminates waveform misalignment caused by differences in sensor response speed, ensuring the accuracy of causal relationships in subsequent feature extraction.
[0056] S1.4: The multi-sensor spatial coordinate mapping matrix is used to perform spatial registration and noise suppression filtering on the dual-channel synchronous mechanical signal pair to eliminate phase distortion introduced by installation position deviation and generate a denoised standardized interface mechanical signal stream.
[0057] The preset multi-sensor spatial coordinate mapping matrix is invoked. This matrix represents the relative positional relationship and installation deviation vector of the micro shear stress sensor array and the high frame rate micro vibration probe in the three-dimensional physical space of the weighing platform.
[0058] Spatial registration is performed on the dual-channel synchronous mechanical signal pair. The signal propagation delay compensation is calculated based on the physical coordinates of the sensor to eliminate phase distortion caused by differences in installation position and generate a spatially aligned mechanical signal sequence.
[0059] An adaptive Wiener filtering algorithm is used to suppress noise in the spatially aligned signal. An optimal linear filter is constructed based on the signal power spectral density estimation to separate the background mechanical vibration noise from the effective mechanical response caused by powder flow.
[0060] The filtered signal is standardized by subtracting the noise mean estimate and then dividing by the noise standard deviation estimate.
[0061] The amplitude normalization process is performed on the standardized shear stress signal and the micro-amplitude resonance signal to unify the dimensions to the [-1,1] interval, thereby eliminating the data scale imbalance caused by the difference in sensitivity of different sensors.
[0062] The normalized dual-channel signals are recombined according to timestamps to generate a denoised and standardized interface mechanical signal stream, ensuring that the data received by the subsequent feature extraction module has a high signal-to-noise ratio and spatiotemporal consistency.
[0063] By using spatial registration and adaptive filtering, the dual-channel synchronous mechanical signal pair from the previous step is transformed into a denoised standardized interface mechanical signal stream, thereby achieving the expected technical effect of eliminating installation deviation phase distortion and improving signal purity.
[0064] For example, the weighing platform is set to 500mm × 500mm, the shear stress sensor array contains 4 nodes, and micro-vibration probes are arranged at the four corners of the platform. The spatial coordinate mapping matrix records the xyz coordinate deviation of each sensor relative to the center of the platform, with a maximum installation error of ±2mm. Spatial registration is performed on the signal with a sampling rate of 10kHz, and the maximum phase delay compensation is calculated to be 0.2ms. Adaptive Wiener filtering is used, with a noise window length of 50ms, and the estimated noise mean is 0.05N and the standard deviation is 0.02N. Standardization is performed on the original shear stress signal X(t), compressing the amplitude range from [-5N, 5N] to [-1, 1]. After processing, the signal-to-noise ratio is improved from 15dB to 35dB, and the phase distortion error is reduced to within 0.01ms. The generated standardized interface mechanical signal stream is directly used for subsequent multimodal data structure encapsulation.
[0065] S1.5: Perform multimodal data structure encapsulation and metadata annotation on the standardized interface mechanical signal flow to construct a raw mechanical dataset containing spatiotemporal attribute labels.
[0066] Step S2: Based on a pre-built knowledge graph of powder flowability, perform physical property parameter matching on the original mechanical dataset to generate a flow state classification label set representing the dominant mode of the current working condition. Specifically, this includes:
[0067] S2.1: Perform time-frequency domain joint transformation processing on the original mechanics dataset, and use the short-time Fourier transform algorithm to extract the frequency band energy distribution characteristics of the shear stress response signal and the transient impact factor of the micro-amplitude resonance signal to generate a multi-dimensional mechanical feature vector set containing frequency domain statistics and time domain abrupt change points.
[0068] Receive the raw mechanical dataset generated by S1.5, which contains spatiotemporal attribute labels and synchronously aligned discrete shear stress response time series and micro-amplitude resonance signal time series.
[0069] Windowing preprocessing was performed on the shear stress response signal. The Hanning window function was selected to suppress spectral leakage. The window length was set to 256 sampling points and the overlap rate was 50% to ensure that the time-frequency resolution met the requirements for capturing unsteady flow characteristics.
[0070] A short-time Fourier transform is performed on the windowed shear stress signal segment to map the time-domain signal to the time-frequency domain. The complex spectral coefficients of each frequency component are calculated to obtain the frequency band energy distribution of the signal within a specific time window.
[0071] Based on the preset frequency band range of powder characteristics, the frequency band energy distribution is integrated to extract the energy proportion of low frequency band (0-50Hz) representing agglomeration and slip and the energy proportion of high frequency band (50-200Hz) representing particle collision, forming a frequency domain feature subset of shear stress.
[0072] The micro-amplitude resonance signal is preprocessed by Hilbert-Huang transform, and the intrinsic mode function is obtained by empirical mode decomposition. High-frequency components that reflect the instantaneous impact response of the platform are then selected.
[0073] Calculate the instantaneous envelope of high-frequency components, identify peak points in the envelope that exceed three times the baseline noise threshold, statistically analyze the peak density and average amplitude per unit time, and generate a transient impact factor characterizing the impact intensity of powder.
[0074] The frequency domain feature subset of shear stress is concatenated with the micro-amplitude resonant transient impact factor, and Z-score normalization is performed to eliminate dimensional differences, thus constructing a unified multidimensional mechanical feature vector.
[0075] The original mechanics dataset contains synchronously aligned discrete shear stress response time series and micro-amplitude resonance signal time series, as well as nanosecond-level timestamps, sensor spatial coordinate identifiers, and flow state fields to be labeled for each signal, serving as the basic input data for subsequent joint time-frequency domain transformation and material property parameter matching.
[0076] Through the above-mentioned time-frequency domain joint transformation and feature fusion processing, the multimodal raw data of the previous step is transformed into a multidimensional mechanical feature vector set containing frequency band energy distribution and transient impact factor, realizing the quantitative characterization of powder mechanical behavior under unsteady flow state, and providing highly discriminative input features for subsequent knowledge graph matching.
[0077] For example, a shear stress signal with a sampling rate of 1 kHz is framed using a Hanning window with a window length of 256 points and an overlap of 128 points. After performing a short-time Fourier transform, the energy E in the 0-50 Hz frequency band is calculated. low Energy E in the 50-200Hz frequency band high If at some moment E low =0.85J, E high =0.15J, then the low-frequency energy proportion is 0.85. EMD decomposition is performed on the micro-amplitude resonance signal to extract the IMF1 component. Its Hilbert envelope is calculated, and 15 peaks exceeding the threshold of 0.5V are detected within 1 second, with an average amplitude of 0.8V. Therefore, the transient impact factor is set to 15*0.8=12. Normalizing [0.85, 0.15, 12] forms a three-dimensional mechanical feature vector, significantly improving the feature sensitivity for flow state identification.
[0078] S2.2: Based on the multidimensional mechanical feature vector set, the powder flowability knowledge graph retrieval engine is called. The cosine similarity matching algorithm is used to associate and map the frequency band energy distribution features and transient impact factors in the feature vectors with the basic physical property parameters such as the angle of repose and the Cartesian index stored in the graph to generate a candidate physical property parameter matching sequence.
[0079] Based on the multidimensional mechanical feature vector set generated by S2.1, the powder flowability knowledge graph retrieval engine is invoked to perform feature space mapping. The frequency band energy distribution components and transient impact factors in the feature vectors are extracted as query keys to construct a high-dimensional retrieval vector.
[0080] It should be noted that the powder flowability knowledge graph is a pre-built structured knowledge base. It uses powder property parameters as nodes and the relationships between properties as edges. It stores standard mechanical response template vectors corresponding to the angle of repose, Cartesian index, internal friction angle, and different powder types. It is used to support the association mapping and matching retrieval between real-time operating conditions and static property parameters.
[0081] Read the basic physical property parameter nodes such as the angle of repose, Cartesian index, and internal friction angle stored in the knowledge graph, and convert the pre-set standard mechanical response templates of each physical property parameter into reference feature vectors.
[0082] The cosine similarity algorithm is used to calculate the cosine value of the angle between the query key value and each reference feature vector, thereby quantifying the geometric similarity between the real-time operating condition signal and the standard physical property template.
[0083]
[0084] in, For real-time multidimensional mechanical feature vectors, For the first The reference eigenvectors corresponding to each physical property parameter and The vectors at the th Dimensional component values.
[0085] Set a similarity threshold range, filter out physical property parameter nodes with cosine similarity greater than the preset threshold, and form a preliminary matching candidate set.
[0086] The physical property parameters in the candidate set are sorted in descending order of similarity score to generate a candidate physical property parameter matching sequence containing the physical property parameter identifier and its corresponding similarity weight.
[0087] By using cosine similarity matching, the multidimensional mechanical features from the previous step are transformed into a sequence of candidate physical property parameters with physical semantic association, realizing a preliminary association mapping from abstract signals to specific powder properties, and providing structured input for subsequent probabilistic reasoning.
[0088] For example, for edible flour materials with a particle size distribution of 100-200 mesh, the multidimensional mechanical feature vector V output by S2.1 includes 32-dimensional frequency band energy and 4-dimensional impact factor. The knowledge graph contains three standard template vectors R1, R2, and R3: "high flowability," "moderate agglomeration," and "strong viscosity." The calculated values are Sim(V,R1)=0.85, Sim(V,R2)=0.92, and Sim(V,R3)=0.45. A threshold of 0.80 is set, and R1 and R2 are selected. A sequence [(R2,0.92),(R1,0.85)] is generated based on the scores, clearly indicating that the current working condition mainly matches the moderate agglomeration characteristic, significantly improving the accuracy of property matching.
[0089] S2.3: Perform Bayesian probabilistic inference calculations on the candidate physical property parameter matching sequences, adjust the matching weights by combining the current environmental temperature and humidity compensation coefficients, quantify the confidence scores of each candidate and the real-time operating conditions, and generate a preliminary judgment result set of the dominant flow mode with probabilistic confidence.
[0090] Receive candidate property parameter matching sequences and real-time environmental temperature and humidity sensor data, and construct a Bayesian inference framework to quantify the confidence level of the flow state.
[0091] Based on the mechanism of powder hygroscopic agglomeration, a nonlinear correction model for the Karl Fischer index and angle of repose based on ambient temperature and humidity is established to calculate the drift compensation coefficient of physical property parameters under the current working conditions.
[0092] A likelihood function is constructed using a Gaussian kernel function, and a probability density mapping is performed between the multidimensional mechanical feature vector and the compensated and corrected physical parameters of the knowledge graph to generate likelihood estimates for each candidate flow state.
[0093] We introduce the historical state transition matrix as the prior probability distribution, and combine it with the likelihood estimate at the current moment to calculate the posterior probability distribution using Bayes' theorem.
[0094]
[0095] in, For the first The posterior probability of a flow state. For observation data In state The likelihood value below, This represents the prior probability.
[0096] The calculated posterior probability distribution is normalized, and interference terms with probability values below a preset noise threshold are removed, while the candidate set of dominant flow patterns with high confidence is retained.
[0097] By using Bayesian probabilistic inference and environmental compensation correction, static property matching is transformed into dynamic probability confidence scoring, enabling robust initial judgment of error characteristics under unsteady flow conditions.
[0098] For example, in a dynamic weighing scenario of flour with a moisture content of 3.5%, the ambient temperature is 25℃ and the relative humidity is 60%. The system retrieves the basic Cartesian index from the knowledge graph, obtaining a base Cartesian index of 45 and an angle of repose of 32°. Based on the temperature and humidity correction model, the humidity compensation coefficient is calculated to be 1.12, the temperature compensation coefficient to be 0.98, the corrected Cartesian index to be 50.4, and the angle of repose to be 31.36°. The real-time mechanical feature vector is extracted, and its likelihood with the "intermittent bridging" state is calculated to be 0.85, and its likelihood with the "stable filling" state to be 0.15. Combining the prior probabilities (bridging state 0.3, stable state 0.7), the posterior probabilities are calculated using the Bayesian formula, yielding a probability confidence score of 0.708 for the "intermittent bridging" state and 0.292 for "stable filling". The system determines the current dominant mode as "intermittent bridging" and outputs the label and its confidence score, effectively eliminating misjudgments caused by environmental fluctuations and significantly improving the accuracy of state recognition.
[0099] S2.4: Based on the initial judgment result set of the dominant flow mode, the state boundary constraint rule library is applied. Low confidence interference terms are eliminated through the maximum a posteriori probability decision logic, and the historical state transition matrix is fused for time-series smoothing filtering to generate the final flow state classification label set representing the dominant mode of the current working condition.
[0100] Receive the initial judgment result set of the dominant flow mode with probability confidence generated by S2.3. The result set includes the posterior probability distribution of each candidate flow state at the current time and the corresponding physical property parameter matching score.
[0101] The system calls a pre-defined state boundary constraint rule library to extract physical state transition constraints compatible with the dynamic weighing scenario of food powder, including minimum dwell time threshold, state mutual exclusion logic matrix, and maximum jump step size limit.
[0102] The low-confidence interference items in the initial judgment result set are filtered by using a state mutual exclusion logic matrix to remove transient change labels that violate physical common sense, such as illegal paths that jump directly from a stable filling state to a high-speed dispersion state.
[0103] Construct a historical state transition matrix based on a Markov chain, count the transition frequency between each flow state in the past N sampling periods, and calculate the state transition probability matrix P, where the elements P ij This represents the probability of transitioning from state i to state j.
[0104] By applying the maximum a posteriori probability decision logic and combining the observation probability at the current moment with the historical transition probability, the corrected state confidence score is calculated using the following formula:
[0105]
[0106] in, For a moment state The corrected confidence level, For the probability of observation, Let be the state transition probability.
[0107] The calculated corrected confidence score is subjected to Viterbi algorithm time-series smoothing filtering, and the optimal state path is traced back to eliminate short-term state jitter caused by sensor noise.
[0108] The state label with the highest modified confidence score and that satisfies the minimum dwell time constraint is selected as the final flow state classification label at the current moment.
[0109] By using maximum a posteriori probability decision-making and temporal smoothing filtering, the initial judgment result of the previous step is transformed into a final flow state classification label set with temporal continuity and physical consistency, thereby significantly improving the robustness of error identification.
[0110] For example, the sampling frequency is set to 1kHz, and the historical window length N is 50. The initial judgment results show a probability of 0.6 for air-solid coupling oscillation and a probability of 0.35 for intermittent bridging. The state transition matrix shows that the previous time step was stable filling, with a probability of 0.1 for transitioning to air-solid coupling oscillation and 0.05 for transitioning to intermittent bridging. Substituting into the formula, the correction score for air-solid coupling oscillation is 0.06, and the correction score for intermittent bridging is 0.0175. After smoothing using the Viterbi algorithm, air-solid coupling oscillation is confirmed as the optimal path. The output label is air-solid coupling oscillation, effectively suppressing misjudgments and significantly improving the accuracy of state recognition.
[0111] Step S3: Based on the flow state classification label set, the waveform segmentation engine is invoked to semantically segment the original mechanical dataset using physical inflection points in the pressure waveform as boundaries, generating a sequence of pressure waveform segmentation units with clear physical meaning. Specifically, this includes:
[0112] S3.1: Based on the flow state classification label set, retrieve the preset segmentation rule base, obtain the dynamic threshold parameter set and inflection point identification logic configuration that match the current dominant flow mode, so as to generate a waveform segmentation control instruction set adapted to the current working condition.
[0113] It should be noted that the waveform segmentation engine is a signal processing function module implemented based on software programs. Its input is the original mechanical dataset and the flow state classification label set. Internally, it encapsulates a segmentation rule base and inflection point recognition logic configuration bound to various flow states, and the output is a sequence of segmented units with clear physical boundaries.
[0114] The flow state classification label set output by S2 is received and used as the configuration index for the waveform segmentation engine.
[0115] Parse the current dominant flow pattern label, perform key-value retrieval in the pre-set segmentation rule base, and locate the inflection point identification logic configuration table bound to this physical mechanism.
[0116] Extract a set of dynamic threshold parameters that match the current operating conditions, including the sensitivity coefficient for sudden pressure drop, the lower limit of slope reversal duration, and the critical value of curvature acceleration.
[0117] The dynamic threshold parameter set is adaptively scaled and corrected based on the differences in powder properties, and the time window length is adjusted by weighting with the Karl exponent to adapt to the response hysteresis characteristics of powders with different flowability.
[0118] The modified dynamic threshold parameter set and inflection point identification logic configuration are encapsulated into a structured instruction object, generating a waveform segmentation control instruction set containing sampling frequency synchronization markers.
[0119] By using the above processing method, the abstract flow state label is transformed into a specific combination of signal processing parameters, realizing a precise mapping from semantic classification to quantitative segmented control, and ensuring that subsequent waveform cutting has clear physical interpretability and adaptability to operating conditions.
[0120] For example, when the flow state label is identified as "gas-solid coupling oscillation," the system retrieves the corresponding segmented control parameters from the rule base. The pressure amplitude drop sensitivity coefficient is set to 0.85, the lower limit of the slope reversal duration is 12ms, and the critical value of curvature acceleration is 1.5Pa / ms². For high Karl Fischer index powders, the time window length is extended from the default 10ms to 15ms to compensate for signal hysteresis. The generated waveform segmented control instruction set includes specific operator types and parameter values, significantly improving the adaptability of feature extraction to unsteady flows and effectively avoiding missegmentation caused by fixed thresholds.
[0121] S3.2: Utilize the waveform segmentation control instruction set to perform multi-scale sliding differential operations on the original mechanical dataset, calculate the first-order slope change rate and second-order curvature acceleration of the pressure signal, and generate an instantaneous derivative feature sequence that characterizes the abrupt change trend of the waveform morphology.
[0122] It receives the waveform segmentation control instruction set generated by S3.1 and the original mechanical dataset output by S1 as the input objects for differential operations.
[0123] Based on the sampling frequency parameters in the instruction set, resampling interpolation is performed on the time series of the pressure signal to ensure that the data point spacing meets the requirements of the Nyquist sampling theorem and to eliminate timing jitter errors caused by asynchronous sensor acquisition.
[0124] Savitzky-Golay smoothing filter is used to preprocess the resampled pressure signal. The window length is set to 5 sampling points and the polynomial order is 2. This preserves the high-frequency change characteristics of the signal while suppressing random noise interference.
[0125] Perform the first-order differential operation of the central difference method on the smoothed pressure signal P(t) to calculate the pressure change rate of adjacent time steps, and generate a first-order slope sequence V(t) that characterizes the instantaneous fluctuation of the powder flow velocity.
[0126] Based on the first-order slope sequence V(t), the second-order differential operation of the central difference method is performed again to calculate the acceleration of the slope change, and a second-order curvature sequence A(t) is generated to characterize the intensity of the sudden change in the powder flow state.
[0127] The first-order slope sequence and the second-order curvature sequence are aligned and merged according to the timestamp to construct a dual-channel instantaneous derivative feature sequence containing information from the velocity domain and the acceleration domain.
[0128] By using the above-mentioned multi-scale sliding differential operation processing method, the original pressure waveform of the previous step is transformed into an instantaneous derivative feature sequence reflecting the flow dynamics characteristics, realizing the quantitative characterization of small morphological changes in unsteady flow, and providing a high signal-to-noise ratio discrimination basis for subsequent adaptive extreme value search.
[0129] For example, for a pressure signal with a sampling frequency of 10 kHz, the Savitzky-Golay filter window is set to 5 points. A 10 ms segment of air-solid coupling oscillation signal is smoothed, reducing the peak-to-peak noise level from 0.5 Pa to 0.05 Pa. First-order differentiation yields a slope sequence with a maximum slope of 200 Pa / s; second-order differentiation yields a curvature sequence with a maximum curvature of 5000 Pa / s². This dual-channel sequence clearly marks three curvature extrema, corresponding to the instant of airbag rupture, significantly improving the accuracy of inflection point identification.
[0130] S3.3: Based on the instantaneous derivative feature sequence and combined with the dynamic threshold parameter set, perform adaptive extreme value search processing to identify pressure amplitude drop points or slope reversal points that satisfy specific flow instability criteria, so as to generate a set of candidate physical inflection point coordinates that mark the physical mechanism jump position.
[0131] The instantaneous derivative feature sequence generated by S3.2 and the dynamic threshold parameter set issued by S3.1 are received as the input data source for adaptive extreme value search.
[0132] The instability criterion logic corresponding to the current flow state label is analyzed, and the amplitude drop sensitivity coefficient and slope reversal duration constraint value are extracted from the dynamic threshold parameter set.
[0133] Perform sliding window extremum detection on the instantaneous first derivative sequence to identify candidate abrupt change points where the derivative value changes from positive to negative and the absolute value exceeds the preset sensitivity coefficient.
[0134] For each candidate mutation point, backtrack to check the rate of change of second-order curvature acceleration within its forward time window to verify whether the curvature continuity condition of the physical mechanism jump is satisfied.
[0135] If the rate of change of curvature exceeds the noise floor threshold, the point is determined to be the point of sudden drop in effective pressure amplitude, and its timestamp coordinates and the corresponding instantaneous pressure value are recorded.
[0136] Parallel processing of instantaneous second derivative sequences is used to search for first derivative sign reversal events near the zero-crossing point of curvature in order to capture slope reversal phenomena in powder flow.
[0137] Calculate the difference between the mean derivatives on both sides of the slope reversal point. If the difference is greater than the minimum gradient jump defined in the dynamic threshold parameter set, then mark it as a valid slope reversal point.
[0138] The identified effective pressure amplitude drop points and effective slope reversal points are merged and then deduplicated and sorted according to their chronological order.
[0139] Redundant inflection points with time intervals shorter than the sensor's minimum response dead zone are eliminated to ensure that there are independent physical process intervals between adjacent inflection points.
[0140] Through the above adaptive extreme value search and multiple criterion verification process, the derivative features of the previous step are transformed into a set of candidate physical inflection point coordinates that mark the position of physical mechanism jump, so as to achieve robust anchoring of key waveform features under unsteady flow conditions.
[0141] For example, when the flow state is "gas-solid coupling oscillation", the amplitude drop sensitivity coefficient is set to 1.5 times the standard deviation, and the slope reversal duration is constrained to 4 ms. Detection of the first derivative sequence reveals that at t=102 ms, the derivative abruptly changes from 5.2 Pa / ms to -8.1 Pa / ms, with the absolute value exceeding the threshold of 6.0 Pa / ms. Backtracking and examining the second-order curvature acceleration within 3 ms forward from this point, its rate of change is found to be 120 Pa / ms², far exceeding the noise floor threshold of 20 Pa / ms², confirming it as a valid drop point. Simultaneously, at t=105 ms, a zero-crossing of the second derivative and a sign reversal of the first derivative are detected, with the mean difference between the two derivatives being 9.5 Pa / ms, greater than the minimum gradient jump of 3.0 Pa / ms, confirming it as a valid slope reversal point. After deduplication and sorting, a candidate physical inflection point coordinate set containing data such as (102ms, -8.1Pa / ms) and (105ms, 0Pa / ms²) is generated, which significantly improves the ability of inflection point identification to suppress high-frequency oscillation interference.
[0142] S3.4: Perform time-series boundary cutting operation on the original mechanics dataset based on the candidate physical inflection point coordinate set, remove invalid noise segments and merge excessively short unsteady intervals to generate a pressure waveform segmented unit sequence with clear physical meaning and time continuity.
[0143] Receive the candidate physical inflection point coordinate set generated by S3.3 and the original mechanical dataset as the input reference for temporal boundary cutting.
[0144] A time-axis segmentation mask is constructed based on the candidate physical inflection point coordinate set, which mathematically divides the continuous pressure waveform data stream into several discrete interval segments to be processed.
[0145] Calculate the duration of each segment to be processed and compare it with a preset minimum effective flow characteristic time threshold for filtering.
[0146] Interval segments with a duration lower than the minimum effective flow characteristic time threshold are removed and identified as invalid signal disturbances caused by high-frequency noise from the sensor or mechanical micro-vibrations.
[0147] Identify short time segments that are adjacent and have the same state label in the remaining intervals, and calculate whether their time interval is less than the state merging tolerance window.
[0148] For adjacent short-term segments that meet the merging conditions, perform data splicing operations to eliminate excessive segmentation caused by minor fluctuations and form a coherent non-steady-state interval.
[0149] The merged pressure waveform segment unit sequence is assigned a unique time sequence index identifier and associated with the corresponding flow state classification label to ensure the integrity of the physical meaning.
[0150] By using the aforementioned temporal boundary cutting and noise removal methods, the discrete inflection points of the previous step are transformed into a sequence of pressure waveform segments with clear physical meaning and continuous time, thereby achieving the expected technical effect of robust anchoring of dynamic weighing error characteristics.
[0151] For example, the minimum effective flow feature time threshold is set to 15ms, and the state merging tolerance window is 5ms. If two adjacent "gas-solid coupling oscillation" segments are detected with durations of 8ms and 6ms respectively and an interval of 2ms, the system merges them into a single effective segment of 16ms. If a segment is only 3ms long, it is directly discarded. In the final output segmented unit sequence, each unit contains a start timestamp, an end timestamp, and a dominant flow state label, significantly improving the signal-to-noise ratio of subsequent feature extraction.
[0152] Step S4: For each segment in the pressure waveform segmentation sequence, perform a feature extraction strategy to calculate a multi-dimensional state attribute vector containing the pressure decay rate curvature extremum or the initial unloading slope, and construct an attributed directed graph structure carrying the temporal and mechanical coupling relationship. Specifically, this includes:
[0153] S4.1: Based on the current flow state classification label, call the pre-set exclusive feature extraction rule library to perform physical mechanism matching processing on the target segmented unit in the pressure waveform segmented unit sequence to generate a multi-dimensional state attribute extraction strategy set containing the pressure decay rate curvature extreme value or the initial unloading slope definition.
[0154] It should be noted that the feature extraction strategy is a set of predefined feature calculation rules for a specific flow state, including feature operator types, threshold parameters and time constraints. After retrieval and matching, the dedicated feature extraction rule library encapsulates it into a multi-dimensional state attribute extraction strategy set, which is used to guide the time-domain differentiation and frequency-domain energy distribution calculation of waveform segmentation units.
[0155] The pressure waveform segmentation unit sequence and the corresponding flow state classification label set output by S3 are received as the input data source for feature extraction strategy matching.
[0156] The flow state label of the current target segment unit is parsed, and key-value retrieval is performed in the pre-set exclusive feature extraction rule base to locate the physical mechanism descriptor and feature operator configuration table bound to the flow mode.
[0157] For gas-solid coupling oscillation, the logic for extracting the extreme value of pressure decay rate curvature defined in the rule base is read, and the parameters of second derivative threshold window and local peak search radius are set.
[0158] For intermittent bridging conditions, the initial unloading slope calculation specification defined in the rule base is read to determine the time starting point offset and minimum sampling point constraints for linear regression fitting.
[0159] Construct a structured object of a multidimensional state attribute extraction strategy set. This strategy set is a concrete instantiation of the feature extraction strategy on the current target segmentation unit. Encapsulate the retrieved feature operators, threshold parameters and time constraints into an executable instruction sequence.
[0160] Through the above matching process, the abstract flow state labels are transformed into a set of specific mathematical operation rules, realizing a precise mapping from semantic classification to quantitative feature calculation, and ensuring that subsequent feature extraction has clear physical interpretability.
[0161] S4.2: Utilize a multidimensional state attribute extraction strategy set (i.e., feature extraction strategy instantiated for the current segmented unit) to perform time-domain differentiation and frequency-domain energy distribution calculation on the original mechanical dataset within the target segmented unit, in order to generate a multidimensional state attribute vector characterizing the instantaneous flow instability.
[0162] The system receives the multi-dimensional state attribute extraction strategy set and the corresponding pressure waveform segmentation unit generated by S4.1, which are used as input objects for time-domain differentiation and frequency-domain energy calculation.
[0163] A sliding window discrete differential operation is performed on the standardized interface mechanical signal flow within the target segmented unit, and the first derivative sequence of the pressure signal is calculated using the central difference method to characterize the instantaneous pressure change rate.
[0164] Based on the first derivative sequence, a discrete differentiation operation is performed again to obtain the second derivative sequence of the pressure signal, which is used to quantify the curvature change characteristics of the pressure waveform.
[0165] For gas-solid coupling oscillation, based on the curvature extremum extraction logic in the strategy set, the local extremum points are searched by traversing the second derivative sequence.
[0166] The time interval variation coefficient between adjacent extreme points is calculated as a time-domain characteristic component characterizing the stability of the oscillation frequency.
[0167] Perform a Fast Fourier Transform on the signals within the same segmented unit to map the time-domain signals to the frequency-domain space.
[0168] Based on the frequency band division rules defined in the strategy set, the proportion of low-frequency energy to total energy is calculated to reflect the intensity of low-frequency pulsations caused by powder agglomeration or bridging.
[0169] Calculate the proportion of high-frequency energy to quantify the high-frequency noise components generated by particle impact or airbag rupture.
[0170] For intermittent bridging conditions, extract signal segments within a specific time window after the pressure peak.
[0171] Perform linear regression fitting on the segment and calculate the absolute value of the slope of the fitted line as the initial unloading slope feature.
[0172] The standard deviation of the fitted residuals is calculated as an auxiliary feature to measure the linearity of the unloading process.
[0173] All calculated time-domain statistics and frequency-domain energy indices are normalized.
[0174] Following the pre-defined feature vector structure order, the normalized feature values are concatenated into a multi-dimensional state attribute vector.
[0175] By using the above-mentioned time-domain differentiation and frequency-domain energy distribution calculation methods, the waveform segmentation data from the previous step is transformed into a multi-dimensional state attribute vector characterizing the instantaneous flow instability, thus achieving accurate quantification and structured expression of the key physical characteristics of the pressure waveform under unsteady flow conditions.
[0176] For example, when the flow state is "gas-solid coupling oscillation", the sampling frequency is set to 1000Hz. Central difference is performed on a segmented unit (50 sampling points) with a length of 50ms, yielding 48 first-order derivative values and 46 second-order derivative values. Local extrema of the second derivative are searched; assuming five maxima are found with time indices of 10, 20, 32, 41, and 48, the adjacent intervals are calculated to be 10, 12, 9, and 7, with a mean of 9.5, a standard deviation of 1.80, and a coefficient of variation of 0.1898. An FFT transformation is performed, assuming a total energy of 100 units, with 60 units of low-frequency energy (0-10Hz) accounting for 0.6% and 15 units of high-frequency energy (50-100Hz) accounting for 0.15%. The resulting multidimensional state attribute vector contains components with a coefficient of variation of 0.197, a low-frequency proportion of 0.6%, and a high-frequency proportion of 0.15. This vector accurately captures the irregularity of the oscillation and the characteristics of the energy distribution, significantly improving the sensitivity of the subsequent error identification model to the gas-solid coupling effect.
[0177] S4.3: Based on the temporal sequence of the pressure waveform segmented unit sequence and the mechanical transmission logic between segments, perform causal correlation analysis on adjacent multidimensional state attribute vectors to generate a set of edge connection relationships describing the temporal dependence of mechanical coupling relationships between segmented units.
[0178] The system receives the multidimensional state attribute vector sequence generated by S4.2 and the pressure waveform segmentation unit time sequence index output by S3 as the input data source for causal correlation analysis.
[0179] Extract the boundary timestamps and flow status labels of two adjacent segment units, calculate the time interval difference between the end time of the previous segment and the start time of the next segment, and construct a temporal proximity index.
[0180] Based on the principles of powder rheology, the inertial transmission effect of the end mechanical state of the previous segment on the initial state of the next segment is analyzed, and the mechanical coupling strength coefficient is defined.
[0181] The cross-correlation function is used to calculate the peak similarity of adjacent segment attribute vectors in the time domain, quantifying the degree of continuity between the two waveforms in terms of physical mechanism.
[0182] Calculate the cross-correlation coefficients of adjacent segment attribute vectors and combine them with the time interval decay factor to obtain the edge weight coefficients.
[0183] Determine the sign and magnitude of the edge weights. If the weight is higher than a preset threshold, establish a directed connection; otherwise, mark it as weakly coupled or disconnect it.
[0184] Record the directionality of the connection, from the previous time series segment to the next time series segment, reflecting the temporal causality of error propagation.
[0185] The calculated edge weights, direction identifiers, and coupling types are encapsulated as edge attribute data and stored in the edge connection relationship set.
[0186] Through the above causal correlation analysis, discrete segmented attribute vectors are transformed into a set of edge connections with topological structure, realizing the transformation from isolated features to spatiotemporal correlation maps, and significantly improving the error identification model's ability to capture the continuity of unsteady flow.
[0187] For example, in the intermittent bridging scenario following air-structure coupling oscillation, the attenuation constant λ is set to 0.04 / ms. The cross-correlation coefficient between adjacent segments on the pressure curvature extremum dimension is calculated to be 0.85, with a time interval Δt of 12ms. The weight of this edge is calculated to be 0.526. Since 0.526 is greater than the threshold of 0.3, the system establishes a strongly coupled directed edge from the oscillation segment to the bridging segment. This edge attribute accurately reflects the physical process of stress redistribution caused by airbag rupture, which in turn triggers bridging. This allows the subsequent graph neural network to effectively trace the source of error based on this connection, significantly improving the robustness of dynamic weighing verification.
[0188] S4.4: Map the multidimensional state attribute vector to node data and the set of edge connection relationships to edge data. Perform topological fusion processing of nodes and edges through a graph structure assembly algorithm to generate a directed graph structure with attributes carrying temporal and mechanical coupling relationships.
[0189] It should be noted that the attributed directed graph structure uses pressure waveform segmentation units as nodes and directed edges are formed by the temporal sequence and mechanical transmission relationship between segmentation units. Each node contains a multi-dimensional state attribute vector corresponding to the segmentation unit as a node feature, and each directed edge contains a weight coefficient representing the mechanical coupling strength between adjacent segments as an edge attribute.
[0190] It receives the multidimensional state attribute vector set generated by S4.2 and the edge connection relationship set generated by S4.3 as the input data source for graph structure assembly.
[0191] Node serialization mapping is performed on the multidimensional state attribute vector. The unique time-series index identifier of each pressure waveform segment unit is mapped to the ID field of the graph node. The state attribute vector containing physical features such as pressure decay rate, curvature extremum, and initial unloading slope is mapped to the feature matrix of the node, thus constructing a node data object with physical semantics.
[0192] The directionality check and weight initialization process is performed on the set of edge connections. The directed attributes of the edges are determined according to the temporal order between the segmented units. The mechanical transmission influence of the preceding segment on the subsequent segment is quantified into the initial weight value of the edge, and an edge data object describing the temporal dependency relationship is constructed.
[0193] The graph topology fusion algorithm is invoked to establish an association index between node data objects and edge data objects using node ID as the key value, ensuring that each directed edge accurately points to its corresponding source node and target node, and eliminating isolated nodes and dangling edges.
[0194] A connectivity test is performed on the completed preliminary graph structure to verify the path integrity from the starting segment unit to the ending segment unit. If there are breakpoints, the missing virtual edge connections are filled in based on timestamp interpolation to ensure the topological continuity of the graph structure.
[0195] The node feature matrix and edge weight matrix are encapsulated into a standardized sparse adjacency matrix format to generate an attributed directed graph structure that carries temporal evolution logic and mechanical coupling strength.
[0196] Through the above graph structure assembly process, discrete segmented features and linear temporal relationships are transformed into graph data representations with complex topological structures, realizing the structured integration of multi-scale mechanical information in unsteady flow processes and providing standardized input for efficient inference of subsequent graph attention networks.
[0197] For example, for a dynamic weighing process of food powder, a sequence containing 5 segmented units is generated. Node 1 corresponds to the "stable filling" state, with a feature vector of [0.1, 0.05]; Node 2 corresponds to the "gas-solid coupling oscillation", with a feature vector of [0.8, 0.9]. The weight of edge E(1->2) is set to 0.7, indicating that the inertial suppression effect of the preceding stable state on the subsequent oscillation is relatively weak. The system assembles these 5 nodes and the 4 directed edges between them into a graph G=(V,E), where V contains the feature vectors of the 5 nodes and E contains the 4 weighted directed edges. The final output graph structure completely preserves the entire mechanical evolution path of the powder from stability to instability and then to recovery, significantly improving the error identification model's ability to capture transient flow characteristics.
[0198] like Figure 3 As shown, step S5: Input the attributed directed graph structure into the pre-trained state-waveform-error map attention network model, perform error-sensitive feature weighted aggregation operation, and generate an error discrimination vector with state traceability. Specifically, this includes:
[0199] S5.1: The State-Waveform-Error Mapping Graph Attention Network Model initializes the node attribute matrix and edge connection relationships in the attributed directed graph structure. It uses a position embedding algorithm to transform the temporal index of the pressure waveform segmentation unit into a high-dimensional position encoding vector to generate an initial node hidden state representation containing spatial topological information and temporal sequence information.
[0200] It should be noted that the state-waveform-error mapping graph attention network model is a pre-trained graph neural network model. Its input is an attributed directed graph structure, and its output is an error discrimination vector with state traceability. The model contains an input embedding layer, a multi-head graph attention layer, a feature aggregation layer, a global pooling layer, and a fully connected classification layer connected in sequence. The input embedding layer is used to fuse the node features and position codes of the attributed directed graph structure into an initial hidden state representation. The multi-head graph attention layer is used to calculate the mechanical coupling strength weights between nodes and perform weighted aggregation of neighborhood features. The feature aggregation layer is used to fuse local waveform features and global flow state labels. The global pooling layer is used to compress node-level features into graph-level global feature descriptors. The fully connected classification layer is used to map the graph-level global feature descriptors into probability distribution weights for each error cause category.
[0201] Receive the attributed directed graph structure output by S4, extract the node attribute matrix and edge connection set, and use it as the input data source for initialization encoding processing.
[0202] For each pressure waveform segmentation unit node in the graph structure, its time index value and corresponding multidimensional state attribute vector are read to construct the original node feature representation.
[0203] A position embedding algorithm combining sine and cosine functions is used to map discrete temporal indices into continuous high-dimensional position encoding vectors to preserve the relative distance information of segmented units on the time axis.
[0204] The calculated high-dimensional position encoding vector is concatenated or linearly added to the original multi-dimensional state attribute vector to generate an initial node hidden state representation containing spatial topological information and time series information.
[0205] Sparse matrix encoding is performed on edge connections to convert mechanical coupling strength into edge weight coefficients and associate them with the corresponding node pairs, thus improving the topological semantics of the graph structure.
[0206] By using location embedding and feature fusion processing, the structured graph data from the previous step is transformed into an initial node hidden state representation with spatiotemporal awareness, thereby achieving effective initialization and traceability anchoring of error features in the graph neural network.
[0207] For example, the position encoding dimension is set to 64. For the segmented unit with a temporal index of 5, a corresponding 64-dimensional position encoding vector is generated. This 64-dimensional position vector is concatenated with the extracted 12-dimensional physical feature vector (such as curvature extrema, unloading slope, etc.) to form a 76-dimensional initial node hidden state. This processing significantly enhances the model's ability to capture temporal dependencies in unsteady flow processes, ensuring that the subsequent attention mechanism can accurately identify error propagation paths.
[0208] S5.2: Based on the initial node hidden state representation, a multi-head self-attention mechanism is executed for calculation. The dynamic association weight coefficients between nodes are obtained by using the linear transformation operation of the query matrix, key matrix and value matrix to generate a normalized attention score matrix that characterizes the mechanical coupling strength between adjacent segment units.
[0209] The initial node hidden state representation matrix H, which contains spatial topology information and time series information, is received from the output of S5.1 and used as the input data source for the multi-head self-attention mechanism calculation.
[0210] Perform a linear projection transformation on the initial node hidden state representation matrix H, using three independent learnable weight matrices W. Q W K W V Generate the query matrix Q, key matrix K, and value matrix V respectively to achieve decoupled mapping of the feature space.
[0211] The dot product operation is performed using the transpose of the query matrix Q and the key matrix K to calculate the original correlation strength score between nodes, thereby quantifying the potential correlation of different pressure waveform segmentation units in mechanical response.
[0212] A scaling factor is introduced to normalize the original association strength score, preventing the gradient from vanishing due to excessively large dot product results and ensuring the numerical stability of the attention distribution.
[0213] The Softmax function is used to perform a probabilistic transformation on the scaled score matrix to generate a normalized attention score matrix A with a row sum of 1, which explicitly represents the relative weights of the mechanical coupling strength between adjacent segment units.
[0214] By performing a weighted summation operation on the normalized attention score matrix A and the value matrix V, global context information is aggregated to capture the modulation effect of long-distance dependencies on local error characteristics during unsteady flow.
[0215] By using a multi-head parallel computing mechanism, the above process is repeated h times, with each head focusing on different subspace features. Finally, the outputs of each head are spliced together and fused through a linear layer to generate a comprehensive attention feature matrix that characterizes the strength of multi-view mechanical coupling.
[0216] By using a multi-head self-attention mechanism, the initial hidden state of the node in the previous step is transformed into a normalized attention score matrix that characterizes the mechanical coupling strength between adjacent segmented units, thereby achieving accurate modeling of complex temporal dependencies under unsteady flow conditions and dynamic weighting of error-sensitive features.
[0217] For example, let the hidden layer dimension d be set. model The number of heads is 64, the number of multiple heads h is 4, and the dimension d of each head is 4. kThe value is 16. The initial node's hidden state H (of shape N×64) is multiplied by the weight matrix W. Q W K W V (All shapes are 64×16), resulting in Q, K, and V. The attention score is calculated as Score = QK. T / 4. Apply Softmax to the score to obtain the attention matrix A. For example, in the gas-solid coupling oscillation state, the attention weights A of the i-th segment and the j-th segment are... ij The value is significantly higher than other static segments, reflecting the energy transfer effect of oscillation. The final output attention feature matrix effectively captures the waveform distortion correlation caused by powder flow instability, significantly improving the robustness of error identification.
[0218] S5.3: The initial node hidden state representation is weighted and aggregated based on the normalized attention score matrix. A nonlinear activation function is used to fuse the local pressure decay rate curvature extremum and the global flow state label features to generate a high-order context node embedding vector that integrates multi-scale spatiotemporal dependencies.
[0219] The normalized attention score matrix output from S5.2 and the initial node hidden state representation generated from S5.1 are received as the input data source for the weighted aggregation operation.
[0220] Perform dimension transpose and broadcast alignment on the attention score matrix to ensure that the attention weight vector of each node is strictly matched with its corresponding hidden state vector in the feature dimension, and construct the tensor structure required for element-wise multiplication.
[0221] By using the Hadamard Product operator to multiply the normalized attention weights element-wise with the initial node hidden state, a weighted feature intermediate state representing the contribution of neighboring node information is generated, thus realizing information filtering based on the strength of mechanical coupling.
[0222] The weighted feature intermediate states are summed and aggregated along the neighborhood dimension to gather multi-source information from temporally adjacent and mechanically related segmented units, forming a preliminary aggregated vector containing local context dependencies.
[0223] A nonlinear activation function, LeakyReLU, is introduced to perform a mapping transformation on the initial aggregated vector. By setting a negative slope parameter of 0.01, the gradient signal of small error features is preserved, thus eliminating the feature saturation effect caused by linear superposition.
[0224] The global flow state label features are projected through a fully connected layer to the same latent space dimension as the node embedding, and then spliced and fused with the activated aggregated vector to inject macroscopic working condition constraint information.
[0225] A multilayer perceptron is used to nonlinearly reconstruct the spliced fusion vector, and a high-order context node embedding vector that incorporates multi-scale spatiotemporal dependencies is calculated.
[0226] Through the aforementioned weighted aggregation and nonlinear fusion processing, the attention weights and initial states from the previous step are transformed into high-order node embedding data with global working condition perception capabilities, thereby achieving precise enhancement of error features in the spatiotemporal dimension and preservation of state traceability.
[0227] For example, for pressure waveform segmentation under gas-solid coupling oscillation state, let the initial hidden state of the current node i be a 64-dimensional vector h. i The attention weight α of its neighbor node j ij The value is 0.85. Perform the Hadamard product operation, and set h... i The dimension value corresponding to the high-frequency oscillation feature is multiplied by 0.85 to suppress low-correlation static background noise. LeakyReLU activation is applied to the aggregated vector; if a dimension value is -0.05, it is mapped to -0.005 to retain weak precursor signals of airbag rupture. The global label encoding [0,1,0,0,0] representing "air-solid coupling" is projected into a 64-dimensional vector and concatenated. After MLP reconstruction, a 128-dimensional high-order embedding vector is output. This vector significantly enhances the weight of curvature extremum features, enabling subsequent pooling layers to more accurately capture weighing deviations caused by oscillations, and greatly improving error recognition sensitivity.
[0228] S5.4: Perform graph pooling on the embedding vectors of higher-order context nodes, and use a global average pooling algorithm based on state traceability constraints to extract key error-sensitive feature components to generate a graph-level global feature descriptor that compresses redundant information and retains the core error pattern.
[0229] It receives the higher-order context node embedding vector output by S5.3 as the input data source for graph pooling operations.
[0230] The node importance assessment based on state traceability constraints is performed on the embedded vector, and the contribution score of each segment unit in the error causes is calculated using the semantic weights of the flow state labels.
[0231] A global average pooling operator is constructed, and the contribution score is used as a weighting coefficient to perform a weighted summation operation on the embedding vector of the higher-order context node.
[0232] The graph-level global feature descriptor is obtained by weighted summation and averaging of the embedding vectors of higher-order context nodes.
[0233] The weighted summation result is normalized using the L2 norm to eliminate dimensional bias caused by differences in the number of segments, generating a standardized global feature vector.
[0234] By using a global average pooling algorithm based on state traceability constraints, the high-order context node embedding vector from the previous step is transformed into a graph-level global feature descriptor that compresses redundant information while retaining the core error pattern, thus achieving robust extraction and dimensionality compression of features sensitive to unsteady flow errors.
[0235] S5.5: Based on graph-level global feature descriptors, nonlinear mapping transformation is performed through fully connected layers. The soft maximum classifier is used to output the probability distribution weights of the error causes in each dimension, so as to generate an error discrimination vector that can clearly indicate the source of unsteady flow error and has state traceability.
[0236] The graph-level global feature descriptor output from S5.4 is received as the input data source for generating the error discrimination vector.
[0237] A fully connected layer nonlinear mapping transformation is performed on the graph-level global feature descriptor. The high-dimensional topological features are projected onto the error cause space using the weight matrix and bias vector to generate the hidden state vector of the intermediate layer.
[0238] By applying the ReLU activation function to the hidden state vector of the intermediate layer for nonlinear rectification, negative noise interference is suppressed and feature sparsity is enhanced, forming an error feature representation with physical interpretability.
[0239] The error features are input into a soft maximum classifier, and the probability distribution weights of each preset error cause category are calculated through exponential normalization to construct a multidimensional probability output vector.
[0240] The maximum likelihood decision is performed on the probability distribution vector, and the category with the largest probability weight is selected as the dominant error cause label, while the complete probability distribution is preserved to support weighted correction in multi-factor coupled scenarios.
[0241] By using fully connected mapping and soft maximum classification, the graph-level global feature descriptor from the previous step is transformed into an error discrimination vector with state traceability, thereby achieving accurate quantification and attribution of the sources of unsteady flow errors.
[0242] For example, when the graph-level global feature descriptor is a 128-dimensional vector, the fully connected layer maps it to a 5-dimensional logits vector, corresponding to five error causes such as air-solid coupling oscillation and intermittent bridging. Assuming the logits value is [2.1, -0.5, 3.4, 1.2, 0.8], after calculation by the soft maximum classifier, the probability of the air-solid coupling oscillation class is 0.68, the probability of the intermittent bridging class is 0.25, and the probabilities of the other classes are low. The system determines that the dominant error is air-solid coupling oscillation and outputs an error discrimination vector containing the probability weights of each type. This vector directly guides the calculation of the compensation factor for air-solid coupling characteristics in step S6, significantly improving the targeting and accuracy of dynamic weighing verification.
[0243] Step S6: Based on the error discrimination vector, analyze the causes of dynamic weighing errors due to unsteady flow, calculate the real-time error correction compensation factor, and output the adaptive verification result for correcting the weighing reading. Specifically, this includes:
[0244] S6.1: Perform semantic mapping processing on the error discrimination vector. Based on the weight distribution of each dimension in the parsed vector of the pre-set state-error attribution rule base, generate an error cause label set containing specific flow instability types to clarify that the dominant error at the current moment comes from specific physical mechanisms such as gas-solid coupling oscillation or intermittent bridging.
[0245] The system receives the state-traceable error discrimination vector output from step S5, which contains the probability distribution weights of error causes in each dimension. Dimensional decoupling is performed on the error discrimination vector to separate independent confidence components corresponding to five dominant flow modes (stable filling, edge collapse, gas-solid coupling oscillation, intermittent bridging, and high-speed dispersion). A pre-built state-error attribution rule base is invoked, which stores the mapping logic between flow instability types and physical error mechanisms. Each independent confidence component is compared with the threshold boundaries in the rule base to identify potential error sources whose confidence exceeds the dynamic judgment threshold. For each identified potential error source, its weight proportion in the error discrimination vector is extracted, and the relative advantage ratio is calculated to exclude secondary interference terms. Using the maximum likelihood estimation principle, the dominant error cause with the highest weight proportion and satisfying the temporal continuity constraint is selected from multiple potential error sources. The selected dominant error cause is mapped to a specific flow instability type label, such as gas-solid coupling oscillation or intermittent bridging. Generate a set of error cause labels containing specific types of flow instability to clarify the physical source of the dominant error at the current moment. Through semantic mapping and rule matching, the results of the previous step are transformed into error cause label set data with clear physical meaning, achieving the expected technical effects of interpretable location of error sources and accurate attribution.
[0246] For example, the error discrimination vector is [0.05, 0.12, 0.78, 0.03, 0.02], corresponding to five flow modes. The state-error attribution rule base sets a dynamic judgment threshold of 0.6. Comparison reveals that the third dimension component, 0.78, exceeds the threshold, identifying it as gas-solid coupling oscillation. The relative dominance ratio is calculated, confirming it as the sole dominant term. An error cause label set {gas-solid coupling oscillation} is generated, clearly indicating that the error originates from the pressure surge caused by the rupture of air pockets within the powder, significantly improving the targeting of subsequent compensation calculations.
[0247] S6.2: Based on the mechanical coupling relationship between the error cause label set and the directed graph structure with attributes, the segmented error quantization model is called to calculate the local deviation amplitude corresponding to each pressure waveform segment unit, so as to generate a local error amplitude sequence that characterizes the deviation of the instantaneous flow state from the ideal weighing trajectory.
[0248] Receive the error cause label set generated by S6.1 and the attributed directed graph structure output by S4, and locate the pressure waveform segment unit node that matches the current dominant flow instability type.
[0249] Extract the node attribute vectors of the target segmented unit from the directed graph structure with attributes, and parse out the instantaneous average pressure, extreme value of pressure change rate, and frequency domain energy distribution coefficient within the time period.
[0250] The preset piecewise error quantization model is invoked, and the corresponding physical deviation calculation kernel function is selected based on the error cause label to establish the mapping relationship between the local pressure response and the ideal static weighing benchmark.
[0251] The difference between the measured pressure value and the ideal static pressure reference value is weighted and summed using a dynamic correction factor to calculate the local deviation amplitude.
[0252] The extracted instantaneous average pressure value is substituted into the measured average pressure value in the unit, and the pressure deviation component is calculated by combining the baseline of the ideal static pressure reference value determined by the corresponding physical property parameters in the powder flowability knowledge graph.
[0253] The kinetic correction coefficient is retrieved based on the error cause label. If the label indicates gas-solid coupling oscillation, the high-frequency proportion of the frequency domain energy distribution coefficient is taken as the input of the kinetic correction coefficient. If the label indicates intermittent bridging, the duration proportion of the extreme value of the pressure change rate is taken as the input of the kinetic correction coefficient.
[0254] A weighted summation operation is performed to obtain a comprehensive local deviation value that includes the effects of dynamic impact and static drift.
[0255] Traverse all nodes in the pressure waveform segmented unit sequence, arrange the local deviation amplitudes of each node in chronological order, and form continuous time series data.
[0256] By using the parallel computation processing of the piecewise error quantification model, the discrete flow state characteristics are transformed into a sequence of local error amplitudes that characterize the deviation of the instantaneous flow state from the ideal weighing trajectory, thereby achieving a refined quantitative description of the weighing error in each minute time period.
[0257] S6.3: Perform time-window integration and inertial smoothing on the local error amplitude sequence, and use an adaptive gain adjustment algorithm to fuse historical error trends and current abrupt change characteristics to generate a real-time error correction compensation factor that can offset the impact of unsteady flow shocks.
[0258] The system receives the local error amplitude sequence output by S6.2, which characterizes the instantaneous deviation of weighing readings under different flow conditions. An adaptive time sliding window is constructed, dynamically adjusting the window length based on the current flow state classification label. The window is shortened under gas-solid coupled oscillation conditions to capture high-frequency abrupt changes, and extended under stable filling conditions to smooth low-frequency noise.
[0259] A trapezoidal numerical integral is performed on the local error amplitude sequence within the sliding window to calculate the cumulative error, eliminate random fluctuation interference, and extract integral characteristic values that reflect the trend of systematic deviation.
[0260] An inertial smoothing filter is introduced, which performs weighted fusion based on the historical value of the compensation factor from the previous moment and the current integral characteristic value, to suppress the drastic jump in the compensation factor caused by the instantaneous impact of powder and ensure the continuity of the correction process.
[0261] An adaptive gain adjustment algorithm is adopted, which dynamically updates the smoothing coefficient according to the confidence weight in the error discrimination vector. When high uncertainty states such as intermittent bridging are detected, the gain is reduced to enhance robustness, and the gain is increased in the steady state to improve the response speed.
[0262] The sliding window integral value and the compensation factor of the previous time step are weighted and fused according to the adaptive gain coefficient to obtain the real-time error correction compensation factor.
[0263] By using the above processing method, the discrete local error sequence is transformed into a continuous and smooth real-time error correction and compensation factor, which effectively offsets the impact of unsteady flow and adaptively improves the weighing accuracy.
[0264] For example, for a powdered food material with an angle of repose of 45 degrees, when a gas-solid coupling oscillation is detected, the sliding window length is set to 50 ms, and the adaptive gain coefficient α(t) is adjusted to 0.3. The collected local error amplitude sequence is integrated to obtain E. int(t) The compensation factor is 0.05N. Assuming the compensation factor C(t-1) was 0.02N at the previous moment, the current compensation factor C(t) is calculated to be 0.029N. This compensation factor effectively smooths out the noise caused by high-frequency oscillations, significantly reducing the standard deviation of the final weighing reading and improving the stability of dynamic measurement.
[0265] S6.4: Based on the real-time error correction compensation factor, perform algebraic superposition calculation on the original weighing sensor readings, execute dynamic zero-point drift correction and sensitivity coefficient reconstruction to generate high-precision adaptive verification results after eliminating unsteady flow interference.
[0266] The generated real-time error correction compensation factor and the original weighing sensor readings are obtained as the input data source for dynamic correction.
[0267] Zero-point drift separation processing is performed on the original weighing sensor readings to extract the baseline offset at the current moment, eliminating static reference deviation caused by long-term powder accumulation or sensor creep.
[0268] Based on the error cause label set to identify the dominant flow state, the corresponding sensitivity reconstruction algorithm is invoked to calculate the instantaneous sensitivity attenuation coefficient under unsteady-state shock.
[0269] The corrected weighing reading is calculated using the following formula:
[0270]
[0271] in, For high-precision adaptive verification results, This is the original weighing sensor reading. This is the zero-point drift baseline offset. This is a compensation factor for real-time error correction. The standard sensitivity coefficient, This is the instantaneous sensitivity attenuation coefficient.
[0272] The calculated zero-point drift baseline offset is subtracted from the original reading, and a real-time error correction compensation factor is added to complete the initial correction at the amplitude level.
[0273] Divide the initially corrected value by the product of the standard sensitivity coefficient and the instantaneous sensitivity attenuation coefficient to complete the nonlinear compensation at the gain level and restore the true mass level.
[0274] The corrected data is subjected to boundary limiting processing to eliminate physically unreachable outliers caused by extreme flow mutations, ensuring the physical rationality of the output results.
[0275] By using algebraic superposition and sensitivity coefficient reconstruction, the results of the previous step are transformed into high-precision adaptive verification data that eliminates unsteady flow interference, thereby significantly improving the dynamic weighing accuracy.
[0276] For example, for flour material with an angle of repose of 45 degrees, under intermittent bridging conditions, the original reading R raw It weighs 500.5g and has a zero-point drift of B. offset The actual weight is 0.2g, the real-time compensation factor Cf is -1.5g, the standard sensitivity S is 1.0, and the instantaneous attenuation coefficient kd is 0.98. Substituting these values into the formula, we get W = (500.5 - 0.2 - 1.5) / (1.0 × 0.98) = 508.98g. This result effectively offsets the negative impact error caused by the bridge collapse. Compared with the uncorrected reading, the measurement accuracy is significantly improved, and the zero-point drift effect is eliminated.
[0277] S6.5: Perform confidence assessment and boundary constraint judgment on the high-precision adaptive verification results. If the verification result exceeds the preset safety threshold, trigger an abnormal alarm signal; otherwise, lock the verification result as the final measurement data and update the state statistical parameters of the powder flowability knowledge graph.
[0278] Receives the high-precision adaptive verification result and corresponding error discrimination vector confidence distribution output by S6.4. Constructs an anomaly detection model based on Mahalanobis distance, and calculates the statistical deviation between the current verification result and standard measurement data under the same flow conditions in the historical stable operating condition benchmark library. Uses a sliding window mechanism to extract the verification residual sequence of the most recent N periods, and calculates the mean vector and covariance matrix of the residuals. Calculates the Mahalanobis distance score of the verification result at the current time using the following formula:
[0279]
[0280] in, Score the Mahalanobis distance. This is the current high-precision adaptive verification result vector. This is the mean vector of similar states in the historical benchmark database. Let covariance matrix be the variance matrix. This indicates the transpose operation.
[0281] The calculated Mahalanobis distance score D is compared with a preset dynamic safety threshold Th. This threshold is dynamically adjusted based on the Cartesian exponent in the powder flowability knowledge graph, with a lower threshold set for high-flowability powders and a higher threshold set for low-flowability powders. If the Mahalanobis distance score D exceeds the dynamic safety threshold Th, the current verification result is deemed unreliable, triggering a system anomaly alarm signal and locking the current weighing data from output. An anomaly log is also recorded for subsequent model iterations. If the Mahalanobis distance score D does not exceed the dynamic safety threshold Th, the current verification result is considered reliable, locked as the final measurement data, and written to the database. The current flow state label, error correction compensation factor, and final measurement deviation value are extracted and used as new observation samples to update the state statistical parameters of the powder flowability knowledge graph. An exponentially weighted moving average algorithm is used to update the baseline mean and covariance matrix of the corresponding flow state in the knowledge graph, achieving online adaptive evolution of model parameters. Through the aforementioned confidence assessment and boundary constraint judgment mechanism, the verification results of the previous step are transformed into final measurement data or abnormal alarm commands with quality certification attributes, thereby realizing the high reliability operation and self-learning capability of the weighing system under non-steady-state flow conditions.
[0282] For example, for flour material with an angle of repose of 35 degrees, under intermittent bridging flow conditions, the sliding window size N is set to 50, and the dynamic safety threshold Th is adjusted to 2.5 based on the Cartesian exponent of 18. The current verification result x = [1000.2g, 0.05N / m²] and the historical baseline mean μ = [1000.0g, 0.04N / m²] are collected, and the diagonal elements of the covariance matrix Σ are [0.01, 0.0001]. The calculated Mahalanobis distance D is 2.24, which is less than the threshold of 2.5, indicating a reliable judgment. This result is locked as the final measurement data, and the mean parameter in the knowledge graph is updated with a learning rate of 0.1, significantly improving the system's recognition accuracy for subsequent similar working conditions.
[0283] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.
[0284] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.
[0285] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for dynamic weighing metrological verification of food powders based on pressure waveform analysis, characterized by, Includes the following steps: The micro-shear stress response signal and micro-amplitude resonance signal of the load-bearing area of the weighing platform are obtained and combined into a raw mechanical dataset; Based on a pre-set knowledge graph of powder flowability, the physical property parameters of the original mechanical dataset are matched to generate a set of flow state classification labels for the dominant mode of the current working condition. Based on the flow state classification label set, the waveform segmentation engine is invoked to semantically segment the original mechanical dataset using the physical inflection points in the pressure waveform as boundaries, generating a sequence of pressure waveform segmentation units with clear physical meaning. For each segment in the pressure waveform segmentation unit sequence, feature extraction is performed, a multi-dimensional state attribute vector including the pressure decay rate curvature extremum or the initial unloading slope is calculated, and an attributed directed graph structure carrying the temporal and mechanical coupling relationship is constructed. The attributed directed graph structure is input into a pre-trained state-waveform-error map attention network model, and an error-sensitive feature weighted aggregation operation is performed to generate an error discrimination vector with state traceability. Based on the error discrimination vector, the dynamic weighing error caused by unsteady flow is analyzed, the real-time error correction compensation factor is calculated, and the adaptive verification result for correcting the weighing reading is output.
2. The method for dynamic weighing and measurement verification of food powder based on pressure waveform analysis according to claim 1, characterized in that, Based on the flow state classification label set, the waveform segmentation engine is invoked to semantically segment the original mechanical dataset using physical inflection points in the pressure waveform as boundaries. The specific process includes: Based on the flow state classification label set, a preset segmentation rule base is retrieved to obtain a dynamic threshold parameter set and inflection point identification logic configuration that match the current dominant flow mode, so as to generate a waveform segmentation control instruction set adapted to the current working condition. The waveform segmentation control instruction set is used to perform multi-scale sliding differential operation on the original mechanical dataset to calculate the first-order slope change rate and second-order curvature acceleration of the pressure signal, so as to generate an instantaneous derivative feature sequence characterizing the abrupt trend of waveform morphology. Based on the instantaneous derivative feature sequence and combined with the dynamic threshold parameter set, an adaptive extreme value search process is performed to identify pressure amplitude drop points or slope reversal points that satisfy specific flow instability criteria, so as to generate a set of candidate physical inflection point coordinates that mark the physical mechanism jump position; Based on the candidate physical inflection point coordinate set, a temporal boundary cutting operation is performed on the original mechanical dataset to remove invalid noise segments and merge excessively short unsteady intervals, so as to generate a pressure waveform segmentation unit sequence with clear physical meaning and time continuity.
3. The method for dynamic weighing and measurement verification of food powder based on pressure waveform analysis according to claim 1, characterized in that, The attributed directed graph structure is input into a pre-trained state-waveform-error map attention network model, and an error-sensitive feature weighted aggregation operation is performed to generate an error discrimination vector with state traceability. The specific process includes: The node attribute matrix and edge connection relationship in the directed graph structure with attributes are initialized and encoded. The temporal index of the pressure waveform segmentation unit is transformed into a high-dimensional position encoding vector using the position embedding algorithm to generate an initial node hidden state representation containing spatial topological information and temporal sequence information. Multi-head self-attention mechanism is performed based on the initial node hidden state representation. The dynamic association weight coefficients between nodes are obtained by using linear transformation operations of the query matrix, key matrix and value matrix to generate a normalized attention score matrix that characterizes the mechanical coupling strength between adjacent segment units. The initial node hidden state representation is weighted and aggregated based on the normalized attention score matrix. A nonlinear activation function is used to fuse the local pressure decay rate curvature extremum and the global flow state label features to generate a high-order context node embedding vector that integrates multi-scale spatiotemporal dependencies. Graph pooling is performed on the embedding vectors of higher-order context nodes, and a global average pooling algorithm based on state traceability constraints is used to extract key error-sensitive feature components to generate a graph-level global feature descriptor that compresses redundant information and retains the core error pattern. Based on graph-level global feature descriptors, a nonlinear mapping transformation is performed through a fully connected layer. A soft maximum classifier is used to output the probability distribution weights of the error causes in each dimension, so as to generate an error discrimination vector that can clearly indicate the source of unsteady flow error and has state traceability.
4. The method for dynamic weighing and measurement verification of food powder based on pressure waveform analysis according to claim 1, characterized in that, The shear stress signal is acquired by a shear stress sensor array on a symmetrical weighing platform through high-frequency sampling control, and the micro-amplitude resonance signal is captured by a high-frame-rate micro-vibration probe.
5. The method for dynamic weighing and measurement verification of food powder based on pressure waveform analysis according to claim 1, characterized in that, Before fusing the shear stress signal and the micro-amplitude resonance signal into a raw mechanical dataset, the method further includes: performing nanosecond-level timestamp alignment processing on the shear stress signal and the micro-amplitude resonance signal, and using a spatial coordinate mapping matrix for spatial registration and noise suppression filtering.
6. The method for dynamic weighing and measurement verification of food powder based on pressure waveform analysis according to claim 1, characterized in that, The physical property parameter matching process specifically includes: performing a joint time-frequency domain transformation on the original mechanical dataset to extract the frequency band energy distribution features, and using a cosine similarity matching algorithm to associate and map the frequency band energy distribution features with the repose angle parameters in the spectrum.
7. The method for dynamic weighing and measurement verification of food powder based on pressure waveform analysis according to claim 1, characterized in that, The process of generating the flow state classification label set specifically includes: calculating the confidence score of the candidate physical property through Bayesian probabilistic inference, and performing time-series smoothing filtering in combination with the historical state transition matrix to remove low-confidence interference items.
8. The method for dynamic weighing and measurement verification of food powder based on pressure waveform analysis according to claim 1, characterized in that, The calculation of the multidimensional state attribute vector specifically includes: calling the dedicated feature extraction rule library corresponding to the flow state classification label set, performing time-domain differentiation and frequency-domain energy distribution calculation on the waveform segmentation unit sequence to obtain the multidimensional state attribute vector.
9. The method for dynamic weighing and measurement verification of food powder based on pressure waveform analysis according to claim 1, characterized in that, The construction of the attributed directed graph structure specifically includes: performing causal association analysis on adjacent multidimensional state attribute vectors to generate edge connection relationships, and mapping the multidimensional state attribute vectors to nodes and the edge connection relationships to edges through a graph structure assembly algorithm, performing topology fusion processing to obtain the attributed directed graph structure.
10. The method for dynamic weighing and measurement verification of food powder based on pressure waveform analysis according to claim 1, characterized in that, After outputting the adaptive verification result, the method further includes: evaluating the confidence level of the adaptive verification result; if it passes, locking it as the final measurement data and updating the powder flowability knowledge graph; otherwise, triggering an abnormal alarm.
