A liquid level accurate control method, system and device for a fresh stewed bird's nest filling process

Abstract

The present application belongs to the field of filling control and liquid level detection, and relates to a liquid level accurate control method, system and equipment for a fresh stewed bird's nest filling process, aiming to solve the problem of deviation between static liquid level and filling transient liquid level, and the problem of being unable to compensate for the influence of sedimentation. The present application comprises: obtaining a liquid weight change signal and a receiving stress signal; extracting a high-frequency energy feature from the receiving stress signal, and generating an instantaneous water content by a nonlinear mapping model in combination with the liquid weight change signal; determining a sedimentation characteristic constant according to a low-frequency recovery time constant of the receiving stress signal, the instantaneous water content and the current cumulative volume; correcting the nonlinear mapping model by using a deviation between a filling solid volume and an apparent volume; inputting the instantaneous water content sequence and the sedimentation characteristic constant sequence into a time convolution prediction network when the cumulative filling amount reaches a standard, and predicting a steady-state liquid level; and outputting a filling termination instruction when the predicted steady-state liquid level reaches the midpoint of a qualified liquid level interval. The present application can predict the influence of sedimentation and realize liquid level control after standing.

Description

Technical Field

[0001] This invention belongs to the field of filling process control and liquid level detection technology, specifically involving a method, system and equipment for precise liquid level control in the filling process of fresh stewed bird's nest. Background Technology

[0002] Freshly stewed bird's nest is composed of long, thin strands of bird's nest mixed with rock sugar water. There is no clear interface between the liquid and solid components. At the moment of final filling, the liquid level inside the bottle is not a stable geometric plane, but rather a complex interface consisting of a foam layer, exposed bird's nest strands, and the water phase. Because the settling velocity of the bird's nest strands in the liquid varies significantly depending on their shape, water content, and local concentration, the solids undergo continuous settling and spatial redistribution within seconds to minutes after filling. This causes the liquid volume inside the bottle to be displaced or redistributed, ultimately resulting in a significant difference between the steady-state liquid level after settling and the instantaneous liquid level immediately after filling.

[0003] Currently, the mainstream control method in the industry is the weighing method, which involves setting a target weight and stopping filling when the weight reaches that target. However, for equal weights of bird's nest material, the actual filling volume formed by settling in the container varies greatly due to differences in solid content, strip distribution density, and moisture content. Simply relying on weighing cannot perceive the space occupied by solids in the liquid phase, resulting in inconsistent liquid levels after the same batch of products has settled, severely affecting product appearance consistency and consumer sensory evaluation. Visual methods, which supplement liquid level information using laser ranging or industrial cameras, can only perceive reflected signals from the liquid surface or foam layer, and cannot penetrate turbid liquids to detect the volume space occupied by internal solids, nor can they reflect the true liquid level state formed after the bird's nest strands have settled.

[0004] This creates an industry pain point: at the moment filling is completed, the static liquid level, judged based on weighing or surface visual information, may be acceptable. However, after several minutes of settling, the volume shrinkage caused by solid sedimentation or liquid phase redistribution causes the final steady-state liquid level to deviate from the acceptable range. Current technology cannot predict and compensate for the impact of sedimentation on the liquid level in advance based on real-time sensing signals during the filling process, thus ensuring that the liquid level meets the requirements after settling at the moment filling stops. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, namely the deviation between the liquid level and the transient liquid level during filling, and the inability to compensate for the effects of sedimentation, this invention provides a method, system, and equipment for precise liquid level control during the filling process of fresh stewed bird's nest.

[0006] In a first aspect, this invention proposes a method for precise liquid level control during the filling process of freshly stewed bird's nest, comprising:

[0007] Acquire signals for changes in liquid weight and bearing stress. High-frequency energy features are extracted from the stress signal, and combined with the liquid outflow weight change signal, instantaneous water content is generated through a nonlinear mapping model. During the filling interval, the settling characteristic constant is determined based on the recovery time constant of the stress signal, combined with the instantaneous moisture content and the current cumulative volume; The apparent volume is reconstructed by extracting the solid accumulation height from the material image sequence, and the nonlinear mapping model is corrected by using the deviation between the injected solid volume accumulated by the instantaneous moisture content and the apparent volume. When the current accumulated volume reaches a preset threshold, the recorded instantaneous water content sequence and sedimentation characteristic constant sequence are input into the temporal convolutional prediction network to obtain the steady-state liquid level height prediction value. When the predicted steady-state liquid level reaches the midpoint of the qualified liquid level range, a filling termination command is output.

[0008] Furthermore, the method for obtaining the liquid weight change signal and the bearing stress signal is as follows: The system receives the raw weight increment data stream output from the weighing link, performs recursive average filtering on the raw weight increment data stream to remove pulse spikes caused by the instantaneous opening and closing of the filling valve, and obtains a smooth mass sequence. The smooth mass sequence is then continuously differentially analyzed using timestamps as indices to generate a signal reflecting the change in liquid output weight that reflects the cumulative process of liquid injection within a unit time period. The system synchronously receives multi-channel dynamic stress data streams output from the strain link, performs full-passband conformal processing on the multi-channel dynamic stress data streams, retains the full-band waveform components from transient impact response to quasi-static gradual change trend, separates transient stress clusters induced by alternating impacts of bird's nest silk and liquid on the bearing interface, and gradually changing stress baselines caused by monotonically increasing hydrostatic pressure of liquid column, and weighted superimposes the transient stress clusters and the gradually changing stress baselines to generate the bearing stress signal.

[0009] Furthermore, high-frequency energy features are extracted from the stress signal, and combined with the liquid outflow weight change signal, instantaneous water content is generated through a nonlinear mapping model. The method is as follows: Adaptive wavelet packet decomposition is performed on the stress signal, and the high-frequency sub-node coefficients obtained from the decomposition are reconstructed into a transient impact time-domain waveform. A sliding window is divided along a unified time axis and the energy value of the transient impact time-domain waveform is calculated window by window, and then spliced ​​into a high-frequency energy characteristic curve. The liquid weight change signal is cumulatively differentially processed at equal intervals according to the step size of the sliding window to generate an instantaneous mass flow rate sequence. The instantaneous mass flow rate sequence is then forcibly aligned with the high-frequency energy characteristic curve using a timestamp. The aligned high-frequency energy characteristic curve and the instantaneous mass flow rate sequence are constructed as a dual-channel input tensor and fed into a pre-built nonlinear mapping model. The nonlinear mapping model uses a cross-manifold embedding network as its core to decouple the impact energy component and the mass flow rate component in the latent space and align them with the manifold of the moisture content scale. After point-by-point regression mapping, the instantaneous moisture content that corresponds one-to-one with the current filling instant is output.

[0010] Furthermore, during the filling interval, based on the recovery time constant of the stress signal, combined with the instantaneous moisture content and the current cumulative volume, a settling characteristic constant is determined, specifically including: During the filling interval, the low-frequency component of the received stress signal is separated to extract the quasi-static stress component resulting from the coupling effect of solid sedimentation and liquid hydrostatic pressure. The recovery time constant is determined based on the recovery transition process of the quasi-static stress component from the fluctuating state to the static reference. The instantaneous moisture content at the start of the filling interval, the current cumulative volume up to that time, and the recovery time constant are input together into the sedimentation feature generation logic to generate the sedimentation feature constant.

[0011] Furthermore, determining the recovery time constant specifically includes: The quasi-static stress component is obtained by performing narrowband low-pass filtering on the received stress signal. Mark the start time of the filling interval, extract the waveform of the transition segment from the quasi-static stress component to the static reference at that time, and construct the low-frequency fluctuation attenuation trajectory; Envelope detection is performed on the low-frequency fluctuation attenuation trajectory, and a relaxed attenuation envelope is fitted. The time taken for the envelope amplitude to attenuate to the allowable deviation range of the static reference is determined as the recovery time constant.

[0012] Furthermore, the apparent volume is reconstructed by extracting the solids accumulation height from the material image sequence, and the nonlinear mapping model is corrected using the deviation between the injected solids volume accumulated at the instantaneous moisture content and the apparent volume. Specifically, this includes: Image acquisition is triggered synchronously during the filling interval to obtain a sequence of continuous material images; The container cavity is located and the solid flocs are semantically segmented frame by frame in the material image sequence. The pixel coordinates of the highest point in the solid accumulation area are extracted and converted into time series data of solid accumulation height through spatial scale mapping. The time-series data of the solid material stacking height and the container cavity geometric model are used to perform three-dimensional reconstruction to reconstruct the apparent volume sequence corresponding to each frame. The instantaneous moisture content is integrated over time along the filling interval, and the result is accumulated with the change signal of liquid output weight to obtain the volume of filled solids. The volume deviation signal is generated by subtracting the volume of the injected solids from the end steady-state value of the apparent volume sequence. The volume deviation signal is used as a model correction supervision quantity to apply directional adjustment to the hidden layer weights of the cross-manifold embedded network in the nonlinear mapping model until the volume deviation signal converges to the preset tolerance range after multiple filling cycles, thus completing the self-correction of the nonlinear mapping model.

[0013] Furthermore, when the current accumulated volume reaches a preset threshold, the recorded instantaneous water content sequence and sedimentation characteristic constant sequence are input into a temporal convolutional prediction network to obtain a predicted steady-state liquid level height, specifically including: Along the filling process, the instantaneous moisture content sequence and the sedimentation characteristic constant sequence are aligned point by point according to a unified timestamp to construct a dual-channel temporal input tensor; The temporal convolutional prediction network includes stacked dilated causal convolutional layers and gated linear units. It performs multi-scale causal feature extraction on the dual-channel temporal input tensor, and outputs a steady-state liquid level prediction value corresponding to the current material state at the end of the time series through residual propagation and gating. During the training phase of the temporal convolutional prediction network, the measured liquid level height after standing in a steady state in historical filling batches is used as the supervision label. The network weights are optimized in reverse by the deviation between the predicted steady-state liquid level height and the supervision label until convergence.

[0014] Furthermore, the method for outputting a filling termination command when the predicted steady-state liquid level reaches the midpoint of the qualified liquid level range is as follows: The qualified liquid level range corresponding to the current filling batch is pre-stored, the upper limit and lower limit of the qualified liquid level range are extracted, and the target liquid level reference value is generated by median calculation; The system continuously receives the steady-state liquid level prediction value updated by the temporal convolutional prediction network after each filling interval, and compares the steady-state liquid level prediction value with the target liquid level reference value in real time along the time axis. When the comparison result determines that the predicted steady-state liquid level height falls within the locking window centered on the target liquid level reference value and bounded by a preset tolerance deviation, a filling termination trigger signal is generated. The filling termination trigger signal is synchronously converted into a valve closing command and a control record marker, and the filling termination command is output.

[0015] In a second aspect, the present invention provides a precise liquid level control system for the filling process of freshly stewed bird's nest, used to execute a precise liquid level control method for the filling process of freshly stewed bird's nest, the system comprising: The signal acquisition module is used to acquire the liquid weight change signal and the bearing stress signal; The instantaneous moisture content generation module is used to extract high-frequency energy features from the stress signal and, in combination with the liquid outflow weight change signal, generate instantaneous moisture content through a nonlinear mapping model. The sedimentation characteristic constant determination module is used to determine the sedimentation characteristic constant during filling intervals based on the recovery time constant of the stress signal, combined with the instantaneous moisture content and the current cumulative volume. The model correction module is used to extract the solid accumulation height from the material image sequence to reconstruct the apparent volume, and to correct the nonlinear mapping model by using the deviation between the volume of injected solids accumulated by the instantaneous moisture content and the apparent volume. The steady-state liquid level prediction module is used to input the recorded instantaneous water content sequence and sedimentation characteristic constant sequence into the temporal convolutional prediction network when the current accumulated volume reaches a preset threshold, so as to obtain the predicted steady-state liquid level height. The filling termination command output module is used to output a filling termination command when the steady-state liquid level to be predicted reaches the midpoint of the qualified liquid level range.

[0016] In a third aspect, the present invention provides an apparatus comprising: At least one processor; and a memory communicatively connected to at least one of the processors; The memory stores instructions that can be executed by the processor to implement a method for precise liquid level control during the filling process of freshly stewed bird's nest.

[0017] The beneficial effects of this invention are: This invention integrates information from the weight dimension and the mechanical response dimension by simultaneously acquiring the liquid weight change signal and the stress signal, enabling the system to distinguish between simple mass accumulation and dynamic mechanical response caused by solid impact and accumulation. This solves the problem that traditional weighing methods cannot detect the behavior of solids in the liquid phase.

[0018] This invention extracts high-frequency energy features from the stress signal and generates instantaneous water content using a nonlinear mapping model in conjunction with the liquid weight change signal. This enables real-time, point-by-point analysis of the liquid water content during the filling process, providing dynamic physical property parameters that change with the filling process for subsequent sedimentation prediction, rather than relying on fixed assumed values ​​for each batch.

[0019] This invention quantifies the sedimentation hysteresis characteristics of solids in the liquid phase into a calculable characteristic index by determining the time constant of the recovery to the static reference based on the low-frequency pressure fluctuation of the stress signal during the filling interval, combined with the instantaneous moisture content and the current cumulative volume. This allows the dynamic differences in sedimentation behavior to be captured and calibrated in real time during the filling process.

[0020] This invention reconstructs the apparent volume using material image sequences and corrects the nonlinear mapping model by using the deviation between the volume of filled solids accumulated at instantaneous moisture content and the apparent volume. It introduces visual spatial information to perform closed-loop correction of the mechanical-moisture content mapping relationship, enabling the model to adapt to changes in material morphology in continuous filling batches and maintain the accuracy of moisture content estimation.

[0021] This invention, when the cumulative filling volume reaches a preset threshold, inputs the recorded instantaneous moisture content sequence and sedimentation characteristic constant sequence into a temporal convolutional prediction network to predict the steady-state liquid level after settling. This makes the decision to stop filling based on the prediction of the future settling state, rather than the judgment of the liquid level or weight at the moment of filling, thereby shifting the control target from the end of filling to the end of settling.

[0022] This invention uses the predicted steady-state liquid level reaching the midpoint of the qualified liquid level range as the output condition for the filling termination command, so that the filling stop time is directly aligned with the target value of the final steady-state liquid level, compensating for the liquid level changes caused by solid sedimentation and reducing the dispersion of liquid level after the same batch of products has been left to stand. Attached Figure Description

[0023] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart of a method for precise liquid level control during the filling process of freshly stewed bird's nest according to the present invention; Figure 2 This is a structural diagram of a precise liquid level control system for the filling process of freshly stewed bird's nest according to the present invention; Figure 3 This is a schematic diagram of the structure of a computer system used to implement the methods, systems, and electronic devices of this application. Detailed Implementation

[0024] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] The first embodiment of the present invention proposes a method for precise control of liquid level during the filling process of freshly stewed bird's nest, comprising: Step S10: Obtain the liquid weight change signal and the bearing stress signal; Step S20: Extract high-frequency energy features from the stress signal and combine them with the liquid outflow weight change signal to generate instantaneous water content through a nonlinear mapping model; Step S30: During the filling interval, based on the recovery time constant of the stress signal, combined with the instantaneous moisture content and the current cumulative volume, determine the settling characteristic constant. Step S40: Extract the solid accumulation height from the material image sequence to reconstruct the apparent volume, and use the deviation between the volume of injected solids accumulated by the instantaneous moisture content and the apparent volume to correct the nonlinear mapping model; Step S50: When the current accumulated volume reaches a preset threshold, the recorded instantaneous water content sequence and sedimentation characteristic constant sequence are input into the temporal convolutional prediction network to obtain the steady-state liquid level height prediction value. Step S60: When the predicted steady-state liquid level height reaches the midpoint of the qualified liquid level range, output a filling termination command.

[0027] To more clearly explain the precise liquid level control method for the filling process of freshly stewed bird's nest according to the present invention, the following is in conjunction with... Figure 1 The steps in the embodiments of the present invention are described in detail below: Step S10: Obtain the liquid weight change signal and the bearing stress signal; In this embodiment, step S10 includes: Step S11: Receive the raw weight increment data stream output by the weighing link, perform recursive average filtering on the raw weight increment data stream to remove pulse spikes caused by the instantaneous opening and closing of the filling valve, and obtain a smooth mass sequence. Perform continuous difference on the smooth mass sequence with timestamp as index to generate a liquid weight change signal that reflects the cumulative process of liquid injection within a unit time period. Step S12: Synchronously receive the multi-channel dynamic stress data stream output by the strain link, perform full-passband conformal processing on the multi-channel dynamic stress data stream, retain the full-band waveform components from the transient impact response to the quasi-static gradual change trend, separate the transient stress clusters induced by the alternating impact of bird's nest silk and liquid on the bearing interface, and the gradually changing stress baseline caused by the monotonically increasing hydrostatic pressure of the liquid column, and weighted superimpose the transient stress clusters and the gradually changing stress baseline to generate the bearing stress signal.

[0028] In this embodiment, the key to acquiring the liquid weight change signal lies in accurately extracting dynamic information that characterizes the effective injection volume per unit time from the interference-laden raw weighing data. On actual filling lines, the raw weight increment data stream output by the weighing sensor is not a smooth numerical curve, but rather a discrete sequence superimposed with a large amount of process noise. The main sources of this noise are the mechanical vibrations and hydraulic shocks caused by the instantaneous opening or closing of the filling valve, which contaminate the weight data in the form of pulse-like spikes. If these raw data are directly differentially analyzed to obtain flow information, the spike interference will be drastically amplified, leading to complete distortion of the instantaneous flow estimation.

[0029] Therefore, this embodiment performs recursive averaging filtering on the original weight increment data stream. In this embodiment, the sampling rate of the weighing sensor is set to 200Hz, and the typical operating cycle of the filling valve is approximately 80 to 120ms, with the duration of the pulse spikes it causes typically ranging from 30 to 50ms. To ensure that the time span of the window is sufficient to cover the vibration attenuation process caused by a single opening and closing, the window length of the recursive averaging filter is set to 10 sampling points, corresponding to a 50ms time window. The weight values ​​of the 10 consecutive sampling points included in this window are averaged with equal weights, and the result is used as the smoothed weight value at the end of the window.

[0030] With each new sampling point, the window is pushed forward by one point, continuously generating a smooth mass flow sequence. This window length strikes a balance between smoothing effect and response speed: a 50ms window is sufficient to suppress most spike fluctuations, while keeping the signal delay caused by smoothing within half the window length, i.e., less than 25ms, so as not to significantly lag behind the actual filling flow rate changes. After obtaining the smooth mass flow sequence, this embodiment uses the timestamp of each sampling point as an index to calculate the first-order backward difference of the weight between two adjacent points, i.e., the weight value of the current point minus the weight value of the previous point and then divided by the sampling period, forming a signal reflecting the cumulative progress of liquid injection within each 5ms unit time period. This signal is essentially a discrete approximation of the instantaneous mass flow rate, and its sign and magnitude directly correspond to the start and stop and speed of filling injection. To further eliminate the high-frequency quantization noise that may be introduced by the differential operation, a first-order low-pass digital filter with a cutoff frequency of 5Hz is connected in series after the differential. This cutoff frequency is much higher than the effective bandwidth of the filling mass flow rate change, smoothing out weak spikes without obscuring the actual flow rate dynamic changes.

[0031] Synchronous with the weighing link, this embodiment receives a multi-channel dynamic stress data stream output from the strain link to generate a bearing stress signal. The generation logic of the bearing stress signal is completely different from that of the liquid weight change signal. This is because the physical object sensed by the strain sensor is the mechanical response of the bearing interface of the filling container. Its signal components are much richer than weight changes, covering full-band waveform information from transient impact responses at the millisecond level to quasi-static gradual trends of several seconds or even tens of seconds. In this embodiment, the strain bridge performs four-channel synchronous acquisition at a sampling rate of 1000Hz, corresponding to the stress changes at four symmetrically distributed measurement points at the bottom of the container. The direct impact of the bird's nest strands and liquid on the bottom of the container during the filling and falling process will excite a transient stress cluster containing broadband frequency components of tens to hundreds of Hz. Its waveform is sharp and its duration is usually 5 to 20 ms, containing indirect information about solid content, impact velocity and morphology. At the same time, as the liquid gradually accumulates in the container, the hydrostatic pressure of the liquid column applies a monotonically increasing continuous force to the bearing interface, forming a gradually changing stress baseline with a frequency band concentrated in the range of 0 to 2 Hz.

[0032] These two factors are highly superimposed in both the time and frequency domains. Any form of conventional filtering will irreversibly lose the details of the transient impact, making it impossible to subsequently extract the solid impact characteristics from the stress signal. Therefore, this embodiment implements full-passband conformal processing on the received multi-channel dynamic stress data stream. Specifically, a linear-phase FIR bandpass filter with a passband of 0.1Hz to 500Hz is designed for the original stress sequence of each channel. Its ripple within the passband is less than 0.1dB, and the group delay is strictly constant. The four channels are processed and then averaged to ensure that the time-domain shape of the output waveform is highly consistent with the original physical impact process, without producing waveform distortion. Based on this, signal separation is the core of generating the received stress signal. This embodiment implements a separation algorithm based on a short-time energy threshold on the conformally processed signal: The short-time root mean square (RMS) value of the signal is calculated using a 5ms sliding window. When the energy in a certain window exceeds three times the estimated energy value of the current slowly varying baseline, the waveform within that window is determined to be a transient stress cluster component, which is then separated from the original waveform. The remaining portion is smoothed to form the slowly varying stress baseline. After separation, this embodiment does not simply select one of them as the effective signal, but rather weights and superimposes the transient stress cluster with the slowly varying stress baseline to generate a unified bearing stress signal. In this embodiment, the weight of the transient component is 0.7, and the weight of the baseline component is 0.3. This ratio was determined during the calibration and training phase of the nonlinear mapping model with the minimum RMS error of instantaneous water content prediction as the optimization objective. The superimposed signal not only completely preserves the energy information of the solid impact but also accurately reflects the hydrostatic pressure trend of liquid level accumulation, laying a unified and lossless signal foundation for subsequent extraction of high-frequency energy features and low-frequency recovery time constants.

[0033] Through the above processing, this embodiment successfully transforms the two most readily obtainable physical quantities, weight and strain, into liquid weight change signals and bearing stress signals containing rich process dynamic information. The two signals are strictly synchronized in time and complementary in physical content, together constituting a comprehensive characterization of the instantaneous state of fresh stewed bird's nest filling.

[0034] Step S20: Extract high-frequency energy features from the stress signal and combine them with the liquid outflow weight change signal to generate instantaneous water content through a nonlinear mapping model; In this embodiment, step S20 includes: Step S21: Perform adaptive wavelet packet decomposition on the stress signal, reconstruct the high-frequency sub-node coefficients obtained from the decomposition into a transient impact time-domain waveform, divide the sliding window along a unified time axis and calculate the energy value of the transient impact time-domain waveform window by window, and stitch them together into a high-frequency energy characteristic curve. Step S22: The liquid weight change signal is subjected to equal-interval cumulative difference according to the step size of the sliding window to generate an instantaneous mass flow rate sequence, and the instantaneous mass flow rate sequence is forcibly aligned with the high-frequency energy characteristic curve by timestamp. Step S23: Construct the aligned high-frequency energy characteristic curve and instantaneous mass flow rate sequence into a dual-channel input tensor, and feed it into a pre-built nonlinear mapping model. The nonlinear mapping model uses a cross-manifold embedding network as its core, performs decoupling of the impact energy component and mass flow rate component and manifold alignment to the moisture content scale in the latent space, and outputs the instantaneous moisture content that corresponds one-to-one with the current filling instant through point-by-point regression mapping.

[0035] The method for constructing the nonlinear mapping model is as follows: Multiple sets of calibration samples of bird's nest mixed with rock sugar water with known true moisture content were collected, and the bearing stress signal and the change signal of liquid output weight were recorded synchronously for each calibration sample. The high-frequency energy characteristic curve is extracted from the stress signal of each calibration sample, and the instantaneous mass flow rate sequence is generated from the liquid outflow weight change signal of each calibration sample. The two are combined with the actual water content calibration value to form a ternary calibration dataset. An initial cross-manifold embedding network is constructed, which includes a first manifold coding branch, a second manifold coding branch, and a joint manifold decoder. The joint optimization objective is set as minimizing the decoupling mutual information between the impact energy component and the mass flow rate component in the latent space, and minimizing the water content regression residual. The ternary calibration dataset is divided into a training set and a validation set. The parameters of the initial cross-manifold embedding network are optimized by an alternating iterative method until the joint optimization objective converges to a preset threshold. The converged network parameters are then solidified into the nonlinear mapping model.

[0036] In this embodiment, adaptive wavelet packet decomposition is performed on the bearing stress signal to extract high-frequency energy features that reflect the impact intensity of the solid material. In the bearing stress signal, transient stress clusters induced by the alternating impact of bird's nest fibers and liquid on the bearing interface occupy the mid-to-high frequency band, and their energy intensity is directly modulated by the solid content and impact velocity. To accurately capture this energy change, this embodiment selects the Daubechies 6 wavelet basis, which has tight support and good time-frequency localization capabilities, to perform a 4-layer adaptive wavelet packet decomposition on the bearing stress signal. During the decomposition process, the algorithm automatically selects the path for further decomposition based on the energy distribution of the signal at each node. When the energy of a node coefficient is lower than 10% of the total energy of that layer, the decomposition stops, thus avoiding wasting computational resources in low-energy regions. After decomposition, all high-frequency sub-node coefficients of the fourth layer are taken, corresponding to the 125Hz to 500Hz frequency band. Low-frequency sub-nodes containing the gradually varying stress baseline are discarded, and the selected high-frequency coefficients are reconstructed into a transient impact time-domain waveform through inverse transform. This waveform strips away the static pressure trend, retaining only the rapid fluctuation component caused by the solid impact.

[0037] To quantify the change in impact intensity over time, this embodiment divides the time axis into sliding windows with a length of 100 ms and a step size of 50 ms, corresponding to 100 sampling points and a 50-sampling-point step for the stress signal. Within each window, the sum of squares of the reconstructed waveform is calculated as the high-frequency energy value at that moment. After all windows are calculated, they are spliced ​​together to form a high-frequency energy characteristic curve aligned with the original time axis. The amplitude of the energy curve increases significantly during periods of concentrated solid impact and tends to the background noise level during periods of only liquid injection, thus providing the first dimension of information for subsequent water content mapping.

[0038] Simultaneously, this embodiment synchronously processes the outflow weight change signal to generate a second dimension of information, namely, the instantaneous mass flow rate sequence. The outflow weight change signal has already been obtained in step S10 through differential and low-pass filtering, reflecting the net injected mass within each 5ms unit time period. In order to strictly match the high-frequency energy characteristic curve on the time scale, it is necessary to perform equal-interval cumulative differential on the outflow weight change signal according to the step size of the sliding window. Specifically, using the above-mentioned 50ms step size as the integration interval, the cumulative sum of the mass flow rate within that time period is calculated every 50ms on the outflow weight change signal. The cumulative sum divided by the 50ms duration yields the average instantaneous mass flow rate within that time period, in grams per second.

[0039] Ten instantaneous mass flow rate sampling points can be generated within a 500ms filling cycle, forming an instantaneous mass flow rate sequence. Numerically, this sequence has the exact same time base and sampling rate as the high-frequency energy characteristic curve, so it can be directly timestamped forcibly aligned, that is, the two sequences are paired point by point according to millisecond-level timestamps to form a one-to-one corresponding dual-channel data stream.

[0040] After obtaining the aligned dual-channel data stream, this embodiment feeds it into a pre-built nonlinear mapping model to generate instantaneous water content. The core of the nonlinear mapping model is a cross-manifold embedding network, the construction and training of which will be detailed later. During the inference phase, the network input is a dual-channel input tensor composed of the current high-frequency energy feature value and the instantaneous mass flow rate value, with each sample being a two-dimensional vector. Internally, the first manifold encoding branch maps the impact energy component to a low-dimensional manifold in the latent space, and the second manifold encoding branch maps the mass flow rate component to another low-dimensional manifold in the latent space. Both branches are three-layer fully connected networks, each containing 128 neurons followed by batch normalization and ReLU activation, with the final output being a 16-dimensional latent variable.

[0041] The joint manifold decoder receives a concatenated vector of two 16-dimensional latent variables and performs feature interaction and decoupling through a fusion layer containing an attention mechanism. This fusion layer calculates the cross-attention weights between the two latent variables, suppressing redundant mutual information and achieving effective decoupling of the impact energy component and mass flow rate component in the latent space. Subsequently, the fused features are progressively regressed through two fully connected layers, ultimately outputting a scalar value. This value is activated by a sigmoid function and linearly mapped to a calibrated moisture content range (0% to 100%), serving as the instantaneous moisture content prediction for the current filling instant. During inference, the entire network independently calculates for each aligned sampling point, outputting an instantaneous moisture content sequence corresponding one-to-one with the filling process timeline. This sequence not only captures the macroscopic trend of moisture content changes during batch injection but also reflects the instantaneous fluctuations in a single impact event, providing high temporal resolution physical property parameters for subsequent determination of sedimentation characteristic constants and steady-state liquid level prediction.

[0042] The following section details the construction and training methods of the nonlinear mapping model, which is a crucial prerequisite for ensuring the accuracy of instantaneous moisture content estimation. Model construction consists of four stages: data acquisition, feature extraction, network initialization, and training.

[0043] During the data acquisition phase, multiple sets of calibration samples containing bird's nest and rock sugar water with known true moisture content were prepared. Specifically, freshly stewed bird's nest was used to prepare 13 gradients of mixed samples with moisture content ranging from 30% to 90% by mass ratio. Five independent samples were prepared for each gradient, totaling 65 calibration samples. Each sample underwent simulated filling on a dedicated calibration filling platform. During the process, stress signals and changes in liquid weight were recorded simultaneously, covering the entire filling cycle from start to finish. Immediately after filling, each sample was taken and dried to constant weight at 105 degrees Celsius using a halogen moisture analyzer. The true moisture content calibration value was calculated based on the mass difference before and after drying, accurate to one decimal place.

[0044] During the feature extraction stage, the high-frequency energy feature curve and instantaneous mass flow rate sequence were extracted from the stress signal and liquid weight change signal of each calibration sample according to the methods described in steps S21 and S22. To expand the training data and cover different stages of the filling process, a 200ms segment was extracted from the entire time series of each sample in 50ms increments. Each segment contained four aligned sampling points, and the water content calibration value at the center of the segment was used as the supervision label for that segment. Thus, thousands of triplet data points could be generated from 65 samples. Each triplet contained a 4-dimensional high-frequency energy feature subsequence, a 4-dimensional instantaneous mass flow rate subsequence, and a scalar true water content label, forming a complete calibration dataset. The dataset was randomly divided into a training set and a validation set at a 4:1 ratio to ensure that data points from the same calibration sample would not appear in both sets simultaneously.

[0045] The structure of the initial cross-manifold embedding network has been described in the inference section. During the training phase, a joint optimization objective needs to be set to drive the network to decouple the impact energy component and the mass flow rate component in the latent space and accurately regress the water content. The joint optimization objective consists of two loss functions. The first part is the loss for minimizing the water content regression residual, which directly uses the mean square error to calculate the squared mean of the difference between the network output and the true water content label. The second part is the loss for minimizing the decoupling mutual information, aiming to make the latent variables output by the two manifold encoding branches as independent as possible. Since mutual information is difficult to calculate directly, this embodiment uses an approximate mutual information estimator, specifically: For each training batch, the latent vectors output by the first branch and the second branch are treated as two random variables. An auxiliary discriminant network is used to estimate their Kullback-Leibler divergence. This network is a small, fully connected network whose input is the concatenation of the two latent vectors, and whose output is a scalar. During training, an adversarial approach is employed, minimizing the discriminator's cross-entropy loss and maximizing the generator's cross-entropy loss. This prevents the discriminator from distinguishing whether the two latent variables come from the same data point, thus forcing the mutual information of the latent variables to be minimized. The total loss function is the sum of the moisture content regression loss and the mutual information loss multiplied by a balance coefficient. The balance coefficient is set to 0.1 to decouple the two physical factors as much as possible while ensuring regression accuracy.

[0046] Training employs an alternating iterative approach for parameter optimization. In each training epoch, the parameters of the manifold coding branch and joint decoder are initially fixed, while only the parameters of the mutual information discriminator network are updated. Several steps are performed on the training set to make the discriminator approximate the true mutual information. The discriminator network parameters are then fixed, and the parameters of the manifold coding branch and joint decoder are updated to minimize the total loss function. The optimizer used is AdamW, with an initial learning rate of 0.001, a weight decay coefficient of 0.0001, and a batch size of 128. When the root mean square error of regression on the validation set decreases by less than 0.01% over 10 consecutive epochs and the mutual information loss tends to stabilize, the joint optimization objective is considered converged, and training stops. At this point, the network architecture and the weight parameters of all layers are solidified into a deployable nonlinear mapping model, which can be directly used for real-time inference at the filling site.

[0047] Through the above signal processing procedure and a carefully trained cross-manifold embedding network, this embodiment achieves accurate estimation of the moisture content at every moment during the filling process of fresh stewed bird's nest. The maximum estimation error does not exceed 1.5% on the validation set, providing reliable real-time physical property input for subsequent sedimentation prediction and liquid level control.

[0048] Step S30: During the filling interval, based on the recovery time constant of the stress signal, combined with the instantaneous moisture content and the current cumulative volume, determine the settling characteristic constant. In this embodiment, step S30 includes: Step S31: During the filling interval, perform low-frequency component separation on the bearing stress signal to extract the quasi-static stress component caused by the coupling effect of solid sedimentation and liquid hydrostatic pressure. Step S32: Determine the recovery time constant based on the recovery transition process of the quasi-static stress component from the fluctuating state to the static reference. Step S33: Input the instantaneous moisture content corresponding to the start time of the filling interval, the current cumulative volume up to that time, and the recovery time constant into the sedimentation feature generation logic to generate the sedimentation feature constant.

[0049] Specifically, determining the recovery time constant includes: Step S321: Perform narrowband low-pass filtering on the received stress signal to obtain the quasi-static stress component; Step S322: Mark the start time of the filling interval, extract the transition waveform of the quasi-static stress component from that time to the static reference, and construct the low-frequency fluctuation attenuation trajectory. Step S323: Perform envelope detection on the low-frequency fluctuation attenuation trajectory and fit a relaxed attenuation envelope. The time taken for the envelope amplitude to attenuate to the allowable deviation range of the static reference is determined as the recovery time constant.

[0050] The logic for generating the settlement features is as follows: Step S331: Input the instantaneous water content and current cumulative volume into a pre-built sedimentation feature generator, which converts the water content and volume into an equivalent apparent viscosity factor based on the solid-liquid two-phase rheological relationship. Step S332: Calculate the reciprocal of the recovery time constant as the settlement driving rate factor, form a two-dimensional parameter pair with the equivalent apparent viscosity factor and the settlement driving rate factor, and send it into a pre-constructed fusion mapping network. The fusion mapping network learns the decision boundary line in the two-dimensional parameter space based on the settlement trajectory data of the calibration sample, maps the two-dimensional parameter pair to the settlement hysteresis axis, and outputs a dimensionless settlement characteristic constant.

[0051] This embodiment separates the low-frequency component of the bearing stress signal during the filling interval, extracting the quasi-static stress component resulting from the coupling effect of solid sedimentation and hydrostatic pressure. The bearing stress signal during the filling process contains high-frequency impact components and low-frequency hydrostatic pressure components. During the filling interval, the high-frequency impact disappears due to the cessation of filling, and the remaining stress changes mainly originate from the redistribution of liquid column pressure caused by solid sedimentation and the transition of the suspended system to static accumulation. To accurately separate this low-frequency component, this embodiment employs a narrowband low-pass filter. Specifically, a second-order Butterworth low-pass filter is designed with a cutoff frequency set to 2Hz, corresponding to a time constant of approximately 80ms. This filter can completely preserve the sub-hertz pressure fluctuations reflecting the sedimentation process while filtering out residual trace high-frequency vibration noise. After passing through this filter, the output of the bearing stress signal is the quasi-static stress component. Its waveform typically exhibits a positive offset above the static reference at the beginning of the filling interval, and then gradually declines in the form of an attenuation curve.

[0052] The recovery time constant is determined based on the recovery transition process of the quasi-static stress component. In this embodiment, the start time of the filling interval is marked as the first sampling point where the instantaneous value of the stress signal begins to monotonically decrease relative to the static reference. The static reference value is zero-point calibrated using the stress signal in the empty bottle state before each filling begins, and the average stress within 1 second before filling is taken as the reference. From the start time of the filling interval, a segment of the waveform of the quasi-static stress component is extracted until it recovers to near the reference, constructing a low-frequency fluctuation attenuation trajectory. In actual data, this trajectory is not an ideal first-order exponential decay, but often exhibits a relaxed multi-feature time superposition form due to the non-uniformity of solid accumulation and multi-path sedimentation. To accurately quantify its attenuation rate, this embodiment performs envelope detection on the low-frequency fluctuation attenuation trajectory: first, the absolute value of the attenuation trajectory is taken, and then its upper envelope is extracted through a low-pass filter with a cutoff frequency of 0.5Hz, resulting in a smooth monotonically decaying envelope curve. Subsequently, the least squares method is used to fit the envelope curve. To accurately describe the hindered relaxation behavior during the settlement process, the fitted model uses the stretched exponential function, whose expression is: ; In the formula, For the envelope amplitude over time The function of change is expressed in millivolts; This is the time taken from the start of the filling interval, in seconds; The initial deviation of the envelope curve from the static baseline at the start of the filling interval is expressed in millivolts. Let be the recovery time constant to be determined, in seconds; The tensile index is fixed at 0.7 in this embodiment to describe the relaxation behavior caused by multipath obstruction during settlement. This is a static reference value, and the unit is millivolts.

[0053] After fitting, the time taken for the envelope amplitude to decay to within the allowable deviation range of the static reference is determined as the recovery time constant. This allowable deviation is taken as three times the noise standard deviation of the quasi-static stress component near the static reference, i.e., when the following conditions are met... At that time, among them The noise standard deviation of the quasi-static stress component, in millivolts, is calculated by taking the standard deviation of the stress signal during the static period before filling, corresponding to the duration. This is the recovery time constant, which is generally in the range of 2 to 8 seconds.

[0054] After obtaining the recovery time constant, this embodiment combines the instantaneous moisture content at the start of the filling interval and the current cumulative volume up to that moment, and inputs them into the sedimentation feature generation logic to generate the sedimentation feature constant. The instantaneous moisture content is given by the output value of the nonlinear mapping model in step S20 at the start of the filling interval, and the current cumulative volume is calculated by dividing the cumulative integral of the liquid weight change signal in step S10 by the average density of the material. The average density is taken as 1.08 grams per milliliter based on the ratio of bird's nest to sugar water in the standard formula. The sedimentation feature generation logic is divided into two sub-steps.

[0055] The first sub-step involves inputting the instantaneous moisture content and current cumulative volume into a pre-constructed sedimentation feature generator to generate an equivalent apparent viscosity factor. The sedimentation feature generator is essentially an empirical mapping network based on the rheological relationship between solid and liquid phases. Its construction process involves determining the apparent viscosity of bird's nest suspensions with different moisture contents within a shear rate range of 0.1 to 10 seconds through numerous rheological experiments, while simultaneously recording the volume of the container occupied by the suspension. Experimental data shows that a decrease in moisture content and an increase in solid volume will lead to a power-law increase in the apparent viscosity of the suspension. In this embodiment, a radial basis function interpolation network is trained using these experimental data. The network inputs are the moisture content, expressed as a percentage, and the cumulative volume, expressed in milliliters. The output is the equivalent apparent viscosity measured at 40 degrees Celsius, expressed in Pascals-seconds, with a value range of approximately 0.01 to 2 Pascals-seconds. During deployment, the network is solidified and used as a sedimentation feature generator. It directly outputs the corresponding equivalent apparent viscosity factor based on the real-time instantaneous water content and cumulative volume. The value range is normalized to 0 to 1, where 0 corresponds to the viscosity of pure water and 1 corresponds to the maximum viscosity under the highest solid content state.

[0056] The second sub-step involves calculating the reciprocal of the recovery time constant as the settling driving rate factor. This factor, with dimensions in Hertz, characterizes the driving force intensity of the settling process; a larger value indicates faster settling. The equivalent apparent viscosity factor obtained in the previous step is combined with the settling driving rate factor to form a two-dimensional parameter pair, which is then fed into a pre-constructed fusion mapping network. The fusion mapping network jointly maps the equivalent apparent viscosity factor, representing settling resistance, and the settling driving rate factor, representing driving force intensity, onto a unified settling hysteresis axis. The network's training set is derived from measured calibration data: intermittent filling tests were conducted under various formulations, recording the instantaneous moisture content, cumulative volume, and recovery time constant for each test. The actual settling degree was calibrated by measuring the ratio of the supernatant height to the total liquid height after prolonged settling, and then converted into a dimensionless settling characteristic constant, also ranging from 0 to 1, where 0 indicates almost no settling and 1 indicates completely rapid settling. The network structure is a small multilayer perceptron, containing two input nodes, 16 hidden layer nodes, and one output node. The hidden layers use the hyperbolic tangent activation function, and the output layer uses the sigmoid activation function. Training employs mean squared error loss and the Adam optimizer to learn a decision boundary in the parameter space. This boundary maps regions of high viscosity and low driving rate to low settling constants, and regions of low viscosity and high driving rate to high settling constants. After training, the fusion mapping network is fixed. During real-time inference at the filling site, the equivalent apparent viscosity factor and settling driving rate factor are input into the network, and the network output is the settling characteristic constant corresponding to the current filling interval, accurate to two decimal places.

[0057] Using the above method, this embodiment calculates a settling characteristic constant that reflects the settling tendency of the material in the bottle in real time during each filling interval. This constant takes into account both the viscous resistance determined by the material's own moisture content and volume, and the dynamic driving force of settling revealed by the stress recovery process, providing a key dynamic physical property scalar for subsequent steady-state liquid level prediction.

[0058] Step S40: Extract the solid accumulation height from the material image sequence to reconstruct the apparent volume, and use the deviation between the volume of injected solids accumulated by the instantaneous moisture content and the apparent volume to correct the nonlinear mapping model; In this embodiment, step S40 includes: Step S41: Simultaneously trigger image acquisition during the filling interval to obtain a sequence of material images in consecutive frames; Step S42: Perform container cavity localization and solid floc semantic segmentation on the material image sequence frame by frame, extract the highest point pixel coordinates in the solid accumulation area, and convert them into solid accumulation height time series data through spatial scale mapping; Step S43: Perform three-dimensional reconstruction of the solid material stacking height time series data and the container cavity geometric model to reconstruct the apparent volume sequence corresponding to each frame; Step S44: Integrate the instantaneous moisture content over time along the filling interval and accumulate it with the liquid weight change signal to obtain the volume of filled solids; Step S45: The volume of the injected solids is subtracted from the end steady-state value of the apparent volume sequence to generate a volume deviation signal; Step S46: Use the volume deviation signal as a model correction supervision quantity to apply directional adjustment to the hidden layer weights of the cross-manifold embedded network in the nonlinear mapping model until the volume deviation signal converges to the preset tolerance range after multiple filling cycles, thus completing the self-correction of the nonlinear mapping model.

[0059] In this embodiment, image acquisition is synchronously triggered during the filling interval to obtain a continuous sequence of material images. The filling interval is an ideal window for acquiring images of the material's accumulation state, as the filling action has stopped, surface disturbances and bubbles gradually subside, and solid flocs tend to be relatively stable in the liquid phase, resulting in optimal imaging conditions. In this embodiment, at the start of each filling interval, the control system sends a hardware trigger signal to an industrial camera installed directly above the filling station. The camera resolution is 1920×1080 pixels, the frame rate is set to 30 frames per second, and the exposure time is fixed at 2 milliseconds. Image acquisition continues throughout the entire filling interval, typically between 3 and 10 seconds, thereby acquiring a continuous sequence of material images containing 90 to 300 frames. The container is a cylindrical transparent glass bottle with an outer diameter of 60 mm, an inner diameter of 56 mm, a height of 120 mm, and a wall thickness of 2 mm. To eliminate the influence of ambient light variations, the filling station is equipped with a ring-shaped LED light source with a color temperature of 6500K and an illuminance uniformity better than 95% within the field of view.

[0060] Frame-by-frame localization of the container cavity and semantic segmentation of solid flocs are performed on the material image sequence. The purpose of container cavity localization is to accurately extract the effective internal region of the container in each frame, eliminating interference from bottle wall reflections, labels, and background. This embodiment employs a circular region extraction algorithm based on edge detection and Hough transform: first, the color image is converted to grayscale; edges are extracted using the Canny edge detection operator; then, a circular contour matching the container's inner diameter is searched using Hough circular transform, with the center coordinates and radius determined at sub-pixel precision, forming a mask region for the container cavity. Within the determined cavity region, semantic segmentation of solid flocs is performed. A lightweight U-Net semantic segmentation network is used, with MobileNetV3 as the encoder backbone, pre-trained on a labeled dataset. The labeled dataset contains 2000 images of bird's nest filling containers with different moisture contents and stacking states; each image is manually labeled pixel-by-pixel with distinctions between solid and non-solid regions. The network output is a binary segmentation map with the same resolution as the input image, where white pixels represent solid floc regions and black pixels represent pure liquid regions. The processing time per frame during inference is approximately 15 milliseconds, fully meeting the real-time processing requirement of 30 frames per second.

[0061] After obtaining the solid object segmentation map for each frame, the pixel coordinates of the highest point within the solid object accumulation area are extracted. Within the container cavity mask area of ​​the segmentation map, the highest position of each column of white pixels is counted, and the median of the highest positions in all columns is taken as the highest point of the solid object accumulation. The pixel coordinates of this point are marked as the pixel row and pixel column. This pixel coordinate is converted into time-series data of the solid object accumulation height through spatial scale mapping: with the pixel row at the bottom of the container cavity as the reference zero height and the pixel row at the top of the container cavity as the full bottle reference height, the pixel row coordinates are converted into physical height in millimeters through linear interpolation. In this embodiment, the physical height from the bottom to the top of the container cavity is the bottle height minus the bottle bottom thickness, i.e., 110 millimeters, corresponding to a pixel span of approximately 980 pixels. Therefore, the spatial scale mapping ratio is approximately 0.112 millimeters per pixel.

[0062] After obtaining the time-series data of the solid material stacking height, it is used to perform 3D reconstruction with the container's internal geometric model to reconstruct the apparent volume sequence corresponding to each frame. The container's internal cavity is a standard cylinder with an inner diameter of 56 mm. The stacking morphology of the solid material within the container is approximately an irregular frustum-shaped stack with the bottom as the base and the stacking height as the top. To simplify the complexity of real-time calculations, an integral approximation model is used for apparent volume reconstruction: the container's internal cavity is discretized along the height direction into several thin-layer micro-elements, each layer having a thickness corresponding to the physical height of one pixel. For the current frame at height h, the apparent volume is obtained by summing the volumes of each thin layer from the bottom to height h. Within each thin layer, the cross-sectional area occupied by the solid material is obtained by multiplying the proportion of white pixels in the horizontal cross-section of the segmentation image at that height by the container's cross-sectional area. The container's cross-sectional area is π multiplied by the square of the radius, i.e., 2463 square millimeters. The apparent volume of the solid material in all thin layers is summed to obtain the apparent volume of the current frame, in milliliters. This calculation is repeated for each frame to obtain the apparent volume sequence synchronized with the material image sequence.

[0063] The instantaneous moisture content is integrated over time during the filling interval and accumulated with the liquid weight change signal to obtain the volume of injected solids. The instantaneous moisture content sequence is generated in step S20 and continues to be recorded during the filling interval. The liquid weight change signal is generated in step S10, and its cumulative integral represents the cumulative total injected mass since the start of filling. The method for calculating the volume of injected solids is as follows: the instantaneous moisture content is numerically integrated on the time axis, with the integration interval from the start of filling to the end of the current filling interval. The integration result represents the cumulative proportion factor of solids to the total mass. This proportion factor is multiplied by the cumulative total injected mass to obtain the cumulative total injected solids mass. The total solids mass is divided by the skeleton density of the bird's nest silk solids. The skeleton density of bird's nest silk is experimentally determined to be 1.35 grams per milliliter to obtain the volume of injected solids in milliliters.

[0064] The volume deviation signal is generated by subtracting the volume of injected solids from the steady-state value at the end of the apparent volume sequence. The steady-state value at the end of the apparent volume sequence is the average of the apparent volumes in the last 10 frames before the end of the filling interval, at which point the solids accumulation has tended to stabilize. This average value represents the actual volume of solids in the container as reconstructed visually. The volume deviation signal is the difference between the volume of injected solids and the apparent volume, in milliliters. If the deviation is positive, it indicates that the volume of solids estimated based on mechanical and mass signals is greater than the visual observation value, i.e., the instantaneous moisture content is underestimated, and the moisture content output by the model is too low, requiring upward adjustment; if the deviation is negative, it needs to be downward adjusted. The absolute value of the deviation signal directly reflects the estimation accuracy of the nonlinear mapping model under the current batch material conditions.

[0065] The volume deviation signal is used as a model calibration supervisor to apply targeted adjustments to the hidden layer weights of the cross-manifold embedding network in the nonlinear mapping model. In practice, instead of retraining the entire network, an online fine-tuning strategy is employed to avoid catastrophic forgetting and ensure real-time performance. Fine-tuning is applied only to the last fully connected layer of the joint manifold decoder, which directly maps the fused latent space features to the water content output value. The loss function for fine-tuning is the square of the volume deviation signal, the optimizer is stochastic gradient descent, the learning rate is set low (0.0001), and the momentum is set to 0.9. After each filling cycle, the model is fine-tuned once, with only a single step update to avoid overfitting the current bottle's error. After multiple filling cycles, as material batches change and operating conditions vary, the volume deviation signal gradually converges to a preset tolerance range. This tolerance range is considered to be three times the standard deviation of the volume measurement noise, i.e., ±1.5 ml. When the volume deviation signal falls within the tolerance range over five consecutive filling cycles, the model has completed self-calibration of the material characteristics for that batch. The model parameters remain unchanged during subsequent fillings until the deviation signal exceeds the range again. If the deviation fails to converge within ten consecutive cycles, an early warning is triggered, indicating a possible sensor malfunction or significant deviation from the material formulation, requiring manual intervention.

[0066] Through the above steps, visual spatial information is introduced into the moisture content estimation loop, which originally relied solely on mechanical and mass signals, forming an outer-loop feedback correction mechanism for sensor fusion. The nonlinear mapping model continuously corrects its parameters based on the actual solids accumulation volume during continuous production, maintaining its adaptability to changes in material morphology, batch variations, and long-term drift. This self-correction mechanism requires no offline calibration or manual sampling; it is completed entirely online during idle intervals in the filling process, without additionally impacting production cycle time, thus achieving a balance between model accuracy and long-term stability.

[0067] Step S50: When the current accumulated volume reaches a preset threshold, the recorded instantaneous water content sequence and sedimentation characteristic constant sequence are input into the temporal convolutional prediction network to obtain the steady-state liquid level height prediction value. In this embodiment, step S50 includes: Step S51: Align the instantaneous moisture content sequence and the sedimentation characteristic constant sequence point by point along the filling process according to a unified timestamp to construct a dual-channel temporal input tensor; Step S52, the temporal convolutional prediction network includes stacked dilated causal convolutional layers and gated linear units, performs multi-scale causal feature extraction on the dual-channel temporal input tensor, and outputs a steady-state liquid level prediction value corresponding to the current material state at the end of the time series through residual propagation and gating filtering. Step S53: During the training phase of the temporal convolutional prediction network, the measured liquid level height after standing in a steady state in historical filling batches is used as the supervision label. The network weights are optimized in reverse by the deviation between the predicted steady-state liquid level height and the supervision label until convergence.

[0068] In this embodiment, after filling begins, as the liquid is continuously injected into the container, the instantaneous moisture content and sedimentation characteristic constant time series are continuously recorded and cached. The instantaneous moisture content is output by the nonlinear mapping model in step S20 at each sampling moment, with the sampling interval consistent with the sliding window step size of the high-frequency energy characteristic curve, which is 50ms. The sedimentation characteristic constant is generated at the end of each filling interval in step S30. The filling intervals are distributed according to the operation cycle of the filling valve, typically triggering a brief valve closure every time approximately 30 ml of liquid is injected. A total of 8 to 15 filling intervals are generated throughout the entire filling cycle, corresponding to 8 to 15 sedimentation characteristic constant sampling points. The instantaneous moisture content sequence is a densely sampled continuous curve, while the sedimentation characteristic constant sequence is a sparsely sampled discrete point sequence; the two are not naturally aligned on the time axis.

[0069] To construct a unified temporal input, this embodiment performs point-by-point alignment and interpolation. Using a 50ms uniform time grid of the instantaneous moisture content sequence as a reference, the sedimentation characteristic constant sequence is linearly interpolated on this grid. The interpolated sedimentation characteristic constant retains the value of the most recent intermission during non-filling intervals, and a new calculated value is taken at the moment a filling interval occurs. For each time step, the instantaneous moisture content value and the interpolated sedimentation characteristic constant value are combined into a two-dimensional vector. The two-dimensional vectors of all time steps are stacked in chronological order to form a dual-channel temporal input tensor. The shape of the tensor is the sequence length multiplied by 2. The sequence length dynamically increases with the filling process. When the current accumulated volume accumulated in step S10 reaches a preset threshold, this embodiment sets this threshold to 75% of the target filling volume, i.e., approximately 90 ml. All dual-channel temporal input tensors from the start of filling to the current moment are then extracted as the input to the prediction network. This setting ensures that the prediction is initiated when about a quarter of the fill volume remains, which provides sufficient information about the early stages of the process while allowing room for adjustments to the timing of the filling termination.

[0070] The temporal convolutional prediction network receives the aforementioned dual-channel temporal input tensor and outputs a predicted steady-state liquid level height corresponding to the current material state. The network architecture is designed around three core requirements: first, it strictly adheres to causality, meaning the output at the current moment does not depend on information from future moments, ensuring the feasibility of real-time prediction; second, it possesses multi-scale temporal modeling capabilities, capturing the dynamic changes in sedimentation behavior across scales from seconds to minutes; and third, it maintains stable gradient propagation over long sequences to avoid the degradation problem of deep networks.

[0071] The network consists of four stacked residual blocks, each containing one layer of dilated causal convolution and one layer of gated linear units. The kernel size of the dilated causal convolution is 5, covering the current time step and four previous historical time steps. The dilation coefficient of each residual block increases in powers of 2, namely 1, 2, 4, and 8, so that the receptive field of the network in the fourth layer reaches approximately 160 time steps, corresponding to a historical window of about 8 seconds, which is sufficient to cover the main attenuation process of the sedimentation signal in a single filling interval. Causal convolution is achieved by performing convolution only in the positive direction of the time axis and using zero padding on the left side to ensure that the temporal order is not disrupted. The gated linear units perform gating filtering on the output of the dilated causal convolution. The calculation process is as follows: the convolution output is divided into two parts along the channel dimension. One part is linearly transformed and used as the feature value, and the other part is activated by Sigmoid and used as the gating weight. The two are multiplied element-wise to form the gated feature output. The gating mechanism enables the network to adaptively select key time patterns related to sedimentation and water content changes, suppressing irrelevant fluctuations.

[0072] The input of each residual block is added to the gated output via residual connections. If there is a channel mismatch before the residual connections, 1×1 convolutions are used for dimension alignment. After all residual blocks are stacked, the output tensor of the last time step is taken, compressed into a one-dimensional feature vector by global average pooling, and then fed into a two-layer fully connected head network. The hidden layer contains 64 neurons activated by ReLU, and the output layer is a single-node linear activation, directly outputting the predicted steady-state liquid level height in millimeters. The steady-state liquid level height is defined as the vertical distance from the free surface of the liquid phase inside the container to the bottom of the bottle after filling and standing for 5 minutes. At this point, the settling process is basically complete, and the liquid level is in a stable state.

[0073] The training phase of the temporal convolutional prediction network utilizes supervised learning based on measured data accumulated from historical filling batches. The training data is prepared by collecting complete records of at least 500 filling batches from a continuous production line. Each batch includes the instantaneous moisture content sequence collected throughout the filling process, the sedimentation characteristic constant sequence for each filling interval, and the steady-state liquid level height accurately measured by a laser rangefinder after 5 minutes of settling, with a measurement accuracy of 0.1 mm. The instantaneous moisture content sequence and the interpolated sedimentation characteristic constant sequence for each batch are constructed into unequal-length dual-channel temporal input tensors using the aforementioned method, with the corresponding steady-state liquid level height serving as the supervision label. To improve training efficiency and standardize input size, sequences shorter than the preset maximum length are padded with zeros at the end, and sequences longer than the preset maximum length are truncated at the end. The maximum sequence length is set to 1600 time steps, corresponding to an 80-second filling time. The training and validation sets are randomly divided by batch at a ratio of 9:1.

[0074] During training, the deviation between the output steady-state liquid level prediction value of the temporal convolutional prediction network and the supervision label is used as the loss function, calculated using the root mean square error (RMSE). The optimizer is AdamW, with an initial learning rate of 0.0005, a weight decay coefficient of 0.0001, a batch size of 32, and 200 training epochs. The RMS error on the validation set is monitored during training; early stopping is triggered when the validation error does not reach a new low within 15 consecutive epochs, and the model weights with the smallest validation error are saved as the final deployment model. The trained temporal convolutional prediction network is then embedded and deployed into the inference module of the filling control system.

[0075] During online inference, after each filling interval is completed and the cumulative volume reaches a preset threshold, the system feeds all dual-channel temporal input tensors up to the current moment into the temporal convolutional prediction network. The network completes forward computation within 10 milliseconds and outputs a predicted steady-state liquid level height corresponding to the current material state of the bottle. This predicted value not only reflects the quality and moisture content of the injected liquid but also incorporates the real-time sedimentation dynamics measured during multiple filling intervals. Therefore, it can accurately predict the actual liquid level position of the current batch of material after settling, providing a reliable numerical basis for subsequent filling termination decisions.

[0076] Step S60: When the predicted steady-state liquid level reaches the midpoint of the qualified liquid level range, output a filling termination command.

[0077] In this embodiment, step S60 includes: Step S61: Pre-store the qualified liquid level range corresponding to the current filling batch, extract the upper limit and lower limit of the qualified liquid level range, and generate the target liquid level reference value through median calculation; Step S62: Continuously receive the steady-state liquid level prediction value updated by the temporal convolutional prediction network after each filling interval, and compare the steady-state liquid level prediction value with the target liquid level reference value in real time along the time axis; Step S63: When the comparison result determines that the predicted steady-state liquid level height falls within the locking window centered on the target liquid level reference value and bounded by the preset tolerance deviation, a filling termination trigger signal is generated. Step S64: Synchronously convert the filling termination trigger signal into a valve closing command and a control record marker, and output the filling termination command.

[0078] In this embodiment, the control system pre-stores the qualified liquid level range corresponding to the current filling batch. This qualified liquid level range is defined by the product process specification, and different product specifications correspond to different ranges. Taking a common 45g package of fresh stewed bird's nest as an example, the upper limit of the qualified liquid level range is 85 mm, the lower limit is 75 mm, and the range width is 10 mm. When the filling batch is initialized, the system retrieves the qualified liquid level range corresponding to the current batch from the formula database, extracts the upper and lower limits of the range, and generates a target liquid level reference value through median calculation, which is the arithmetic mean of the upper and lower limits, in this example, 80 mm. The target liquid level reference value means that if the stable liquid level after filling and settling falls exactly at this value, the product's liquid level consistency is optimal, and it provides a symmetrical upper and lower tolerance space for normal fluctuations in production.

[0079] During the filling process, the temporal convolutional prediction network updates its output steady-state liquid level prediction value after each filling interval. This prediction is updated synchronously with the end of the filling interval, and each update is a recalculation based on the complete instantaneous moisture content sequence and sedimentation characteristic constant sequence from the start of filling to the current moment. As the filling volume increases, the prediction value gradually converges to the final steady-state liquid level estimate. The control system continuously receives these updated prediction values ​​and compares each prediction value with the target liquid level reference value in real time along the time axis. The physical meaning of this comparison is to determine whether, given the current filled material, the liquid level will reach the target value after settling if filling is stopped immediately.

[0080] The comparison process does not require the predicted value to be exactly equal to the target liquid level reference value. Instead, a locking window mechanism is used for judgment. The locking window is centered on the target liquid level reference value and bounded by a preset tolerance deviation. The preset tolerance deviation is determined based on the product liquid level tolerance grade and production process capability. In this embodiment, it is half the width of the qualified liquid level range, i.e., 2.5 mm, corresponding to a locking window range of 80 ± 2.5 mm, i.e., from 77.5 mm to 82.5 mm. When the predicted steady-state liquid level height, updated after a filling interval, first falls within this locking window, it is determined that the amount of liquid already filled will reach near the midpoint of the qualified liquid level range after settling. Since the remaining amount that has not yet been filled can still be fine-tuned through subsequent injections, the system generates a filling termination trigger signal at this time. If the predicted value has not yet entered the locking window and is below the lower limit of the window, it indicates that the current filling amount is insufficient, and the system keeps the filling valve open and continues to inject. If the predicted value has exceeded the upper limit of the locking window, it indicates that the filling amount is excessive, but due to the causal constraint, it cannot be withdrawn, and the system still immediately triggers termination to prevent further exceeding of the limit.

[0081] After the filling termination trigger signal is generated, it is synchronously converted into a valve closing command and a control log marker. The valve closing command drives the solenoid valve coil of the filling valve with a 24-volt signal through the digital output module. The solenoid valve completes mechanical closure within 20 milliseconds, cutting off the liquid injection path. Considering the mechanical delay of valve closure and the dripping of residual liquid in the pipeline, the system issues a command when the predicted value is close to the center of the locking window. This advance margin compensates for the extra injection caused by the closing delay, and the magnitude of the advance margin is calibrated to 0.3 ml equivalent mass based on pipeline pressure and valve response time. The control log marker includes the batch number of the current filled bottle, the final filling mass, the instantaneous moisture content at the end of filling, the mean of the previous sedimentation characteristic constant sequence, and the predicted steady-state liquid level height. This marker is written to the filling log database for product quality traceability and subsequent model iteration optimization.

[0082] Through the above steps, the filling termination decision has shifted from the traditional method relying on transient weighing or visual signals to a feedforward control method based on sedimentation prediction. Each filling termination timing is directly aligned with the midpoint of the acceptable liquid level range after settling. This not only specifically compensates for potential level deviations caused by solid sedimentation but also controls the dispersion of liquid levels within a 2.5 mm half-tolerance range for the same batch of products after settling, significantly improving product appearance consistency and filling qualification rate.

[0083] like Figure 2 As shown, a liquid level precision control system for the filling process of fresh stewed bird's nest according to a second embodiment of the present invention is used to execute the above-described liquid level precision control method for the filling process of fresh stewed bird's nest. The system includes: The signal acquisition module is used to acquire the liquid weight change signal and the bearing stress signal; The instantaneous moisture content generation module is used to extract high-frequency energy features from the stress signal and, in combination with the liquid outflow weight change signal, generate instantaneous moisture content through a nonlinear mapping model. The sedimentation characteristic constant determination module is used to determine the sedimentation characteristic constant during filling intervals based on the recovery time constant of the stress signal, combined with the instantaneous moisture content and the current cumulative volume. The model correction module is used to extract the solid accumulation height from the material image sequence to reconstruct the apparent volume, and to correct the nonlinear mapping model by using the deviation between the volume of injected solids accumulated by the instantaneous moisture content and the apparent volume. The steady-state liquid level prediction module is used to input the recorded instantaneous water content sequence and sedimentation characteristic constant sequence into the temporal convolutional prediction network when the current accumulated volume reaches a preset threshold, so as to obtain the predicted steady-state liquid level height. The filling termination command output module is used to output a filling termination command when the steady-state liquid level to be predicted reaches the midpoint of the qualified liquid level range.

[0084] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related explanations of the methods described above can be found in the corresponding processes in the foregoing system embodiments, and will not be repeated here.

[0085] It should be noted that the liquid level precision control system for the fresh stewed bird's nest filling process provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiments can be merged into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the various modules or steps and are not considered as an improper limitation of the present invention.

[0086] A device according to a third embodiment of the present invention includes: At least one processor; and a memory communicatively connected to at least one of the processors; The memory stores instructions that can be executed by the processor to implement the above-described method for precise liquid level control in the filling process of freshly stewed bird's nest.

[0087] A computer-readable storage medium according to a fourth embodiment of the present invention stores computer instructions, which are executed by the computer to implement the above-described method for precise liquid level control in the filling process of freshly stewed bird's nest.

[0088] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the storage device and processing device described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0089] The following is for reference. Figure 3 It shows a schematic diagram of the structure of a computer system for implementing embodiments of the systems, methods, and electronic devices of this application. Figure 3 The server shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0090] like Figure 3As shown, the computer system includes a Central Processing Unit (CPU) 301, which can perform various appropriate actions and processes based on programs stored in Read Only Memory (ROM) 302 or programs loaded from storage section 308 into Random Access Memory (RAM) 303. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An Input / Output (I / O) interface 305 is also connected to the bus 304.

[0091] The following components are connected to I / O interface 305: an input section 306 including a keyboard, mouse, etc.; an output section 307 including a cathode ray tube (CRT), liquid crystal display (LCD), and speakers, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card and a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

[0092] Specifically, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0093] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0094] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0095] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0096] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0097] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for precise liquid level control during the filling process of freshly stewed bird's nest, characterized in that, include: Acquire signals for changes in liquid weight and bearing stress. High-frequency energy features are extracted from the stress signal, and combined with the liquid outflow weight change signal, instantaneous water content is generated through a nonlinear mapping model. During the filling interval, the settling characteristic constant is determined based on the recovery time constant of the stress signal, combined with the instantaneous moisture content and the current cumulative volume; The apparent volume is reconstructed by extracting the solid accumulation height from the material image sequence, and the nonlinear mapping model is corrected by using the deviation between the injected solid volume accumulated by the instantaneous moisture content and the apparent volume. When the current accumulated volume reaches a preset threshold, the recorded instantaneous water content sequence and sedimentation characteristic constant sequence are input into the temporal convolutional prediction network to obtain the steady-state liquid level height prediction value. When the predicted steady-state liquid level reaches the midpoint of the qualified liquid level range, a filling termination command is output.

2. The method for precise liquid level control in the filling process of freshly stewed bird's nest according to claim 1, characterized in that, The method for obtaining the liquid outflow weight change signal and the bearing stress signal is as follows: The system receives the raw weight increment data stream output from the weighing link, performs recursive average filtering on the raw weight increment data stream to remove pulse spikes caused by the instantaneous opening and closing of the filling valve, and obtains a smooth mass sequence. The smooth mass sequence is then continuously differentially analyzed using timestamps as indices to generate a signal reflecting the change in liquid output weight that reflects the cumulative process of liquid injection within a unit time period. The system synchronously receives multi-channel dynamic stress data streams output from the strain link, performs full-passband conformal processing on the multi-channel dynamic stress data streams, retains the full-band waveform components from transient impact response to quasi-static gradual change trend, separates transient stress clusters induced by alternating impacts of bird's nest silk and liquid on the bearing interface, and gradually changing stress baselines caused by monotonically increasing hydrostatic pressure of liquid column, and weighted superimposes the transient stress clusters and the gradually changing stress baselines to generate the bearing stress signal.

3. The method for precise liquid level control in the filling process of freshly stewed bird's nest according to claim 2, characterized in that, High-frequency energy features are extracted from the stress signal, and combined with the liquid outflow weight change signal, instantaneous water content is generated through a nonlinear mapping model. The method is as follows: Adaptive wavelet packet decomposition is performed on the stress signal, and the high-frequency sub-node coefficients obtained from the decomposition are reconstructed into a transient impact time-domain waveform. A sliding window is divided along a unified time axis and the energy value of the transient impact time-domain waveform is calculated window by window, and then spliced ​​into a high-frequency energy characteristic curve. The liquid weight change signal is cumulatively differentially processed at equal intervals according to the step size of the sliding window to generate an instantaneous mass flow rate sequence. The instantaneous mass flow rate sequence is then forcibly aligned with the high-frequency energy characteristic curve using a timestamp. The aligned high-frequency energy characteristic curve and the instantaneous mass flow rate sequence are constructed as a dual-channel input tensor and fed into a pre-built nonlinear mapping model. The nonlinear mapping model uses a cross-manifold embedding network as its core to decouple the impact energy component and the mass flow rate component in the latent space and align them with the manifold of the moisture content scale. After point-by-point regression mapping, the instantaneous moisture content that corresponds one-to-one with the current filling instant is output.

4. The method for precise liquid level control in the filling process of freshly stewed bird's nest according to claim 1, characterized in that, During the filling interval, based on the recovery time constant of the stress signal, combined with the instantaneous moisture content and the current cumulative volume, the settling characteristic constant is determined, specifically including: During the filling interval, the low-frequency component of the received stress signal is separated to extract the quasi-static stress component resulting from the coupling effect of solid sedimentation and liquid hydrostatic pressure. The recovery time constant is determined based on the recovery transition process of the quasi-static stress component from the fluctuating state to the static reference. The instantaneous moisture content at the start of the filling interval, the current cumulative volume up to that time, and the recovery time constant are input together into the sedimentation feature generation logic to generate the sedimentation feature constant.

5. The method for precise liquid level control in the filling process of freshly stewed bird's nest according to claim 4, characterized in that, Determining the recovery time constant specifically includes: The quasi-static stress component is obtained by performing narrowband low-pass filtering on the received stress signal. Mark the start time of the filling interval, extract the waveform of the transition segment from the quasi-static stress component to the static reference at that time, and construct the low-frequency fluctuation attenuation trajectory; Envelope detection is performed on the low-frequency fluctuation attenuation trajectory, and a relaxed attenuation envelope is fitted. The time taken for the envelope amplitude to attenuate to the allowable deviation range of the static reference is determined as the recovery time constant.

6. The method for precise liquid level control in the filling process of freshly stewed bird's nest according to claim 1, characterized in that, The apparent volume is reconstructed by extracting the solids accumulation height from the material image sequence. The nonlinear mapping model is then corrected using the deviation between the injected solids volume accumulated at the instantaneous moisture content and the apparent volume. Specifically, this includes: Image acquisition is triggered synchronously during the filling interval to obtain a sequence of continuous material images; The container cavity is located and the solid flocs are semantically segmented frame by frame in the material image sequence. The pixel coordinates of the highest point in the solid accumulation area are extracted and converted into time series data of solid accumulation height through spatial scale mapping. The time-series data of the solid material stacking height is used to perform three-dimensional reconstruction with the container cavity geometric model to reconstruct the apparent volume sequence corresponding to each frame. The instantaneous moisture content is integrated over time along the filling interval and accumulated with the change signal of liquid output weight to obtain the volume of filled solids; The volume deviation signal is generated by subtracting the volume of the injected solids from the end steady-state value of the apparent volume sequence. The volume deviation signal is used as a model correction supervision quantity to apply directional adjustment to the hidden layer weights of the cross-manifold embedded network in the nonlinear mapping model until the volume deviation signal converges to the preset tolerance range after multiple filling cycles, thus completing the self-correction of the nonlinear mapping model.

7. The method for precise liquid level control in the filling process of freshly stewed bird's nest according to claim 1, characterized in that, When the current accumulated volume reaches a preset threshold, the recorded instantaneous water content sequence and sedimentation characteristic constant sequence are input into a temporal convolutional prediction network to obtain a predicted steady-state liquid level height, specifically including: Along the filling process, the instantaneous moisture content sequence and the sedimentation characteristic constant sequence are aligned point by point according to a unified timestamp to construct a dual-channel temporal input tensor; The temporal convolutional prediction network includes stacked dilated causal convolutional layers and gated linear units. It performs multi-scale causal feature extraction on the dual-channel temporal input tensor, and outputs a steady-state liquid level prediction value corresponding to the current material state at the end of the time series through residual propagation and gating. During the training phase of the temporal convolutional prediction network, the measured liquid level height after standing in a steady state in historical filling batches is used as the supervision label. The network weights are optimized in reverse by the deviation between the predicted steady-state liquid level height and the supervision label until convergence.

8. The method for precise liquid level control in the filling process of freshly stewed bird's nest according to claim 1, characterized in that, The method for outputting a filling termination command when the predicted steady-state liquid level reaches the midpoint of the qualified liquid level range is as follows: The qualified liquid level range corresponding to the current filling batch is pre-stored, the upper limit and lower limit of the qualified liquid level range are extracted, and the target liquid level reference value is generated by median calculation; The system continuously receives the steady-state liquid level prediction value updated by the temporal convolutional prediction network after each filling interval, and compares the steady-state liquid level prediction value with the target liquid level reference value in real time along the time axis. When the comparison result determines that the predicted steady-state liquid level height falls within the locking window centered on the target liquid level reference value and bounded by a preset tolerance deviation, a filling termination trigger signal is generated. The filling termination trigger signal is synchronously converted into a valve closing command and a control record marker, and the filling termination command is output.

9. A precise liquid level control system for the filling process of freshly stewed bird's nest, used to execute the precise liquid level control method for the filling process of freshly stewed bird's nest as described in any one of claims 1-8, characterized in that, The system includes: The signal acquisition module is used to acquire the liquid weight change signal and the bearing stress signal; The instantaneous moisture content generation module is used to extract high-frequency energy features from the stress signal and, in combination with the liquid outflow weight change signal, generate instantaneous moisture content through a nonlinear mapping model. The sedimentation characteristic constant determination module is used to determine the sedimentation characteristic constant during filling intervals based on the recovery time constant of the stress signal, combined with the instantaneous moisture content and the current cumulative volume. The model correction module is used to extract the solid accumulation height from the material image sequence to reconstruct the apparent volume, and to correct the nonlinear mapping model by using the deviation between the volume of injected solids accumulated by the instantaneous moisture content and the apparent volume. The steady-state liquid level prediction module is used to input the recorded instantaneous water content sequence and sedimentation characteristic constant sequence into the temporal convolutional prediction network when the current accumulated volume reaches a preset threshold, so as to obtain the predicted steady-state liquid level height. The filling termination command output module is used to output a filling termination command when the steady-state liquid level to be predicted reaches the midpoint of the qualified liquid level range.

10. A device, characterized in that, include: At least one processor; and a memory communicatively connected to at least one of the processors; The memory stores instructions that can be executed by the processor to implement a precise liquid level control method for the filling process of fresh stewed bird's nest as described in any one of claims 1-8.

Images