Silo material opening discharging device and anti-blocking method thereof based on monitoring and dredging

Abstract

The invention discloses a silo material opening discharging device and an anti-blocking method based on dredging monitoring. The discharging device comprises a discharging barrel, a dredging executing mechanism, an annular sensor matrix and a control center. The anti-blocking method comprises the following steps: acquiring a pressure signal of a flowing material; the edge computing gateway decouples and preprocesses the pressure signal, and outputs a standardized time sequence data sequence; the AI analysis unit is used for extracting spatial features, time domain features and frequency domain features of the multi-dimensional monitoring data, inputting an attention mechanism for fusion, and inputting a fusion feature vector into a blockage prediction model for state reasoning; controlling a hierarchical grooming action according to a reasoning result; according to the anti-blocking method based on monitoring and dredging of the discharging device, sensing, calculation, AI reasoning and intelligent execution are integrated, an intelligent control closed loop is formed, manual experience and intermittent operation can be replaced, and continuous, autonomous and intelligent anti-blocking guarantee in the discharging process is achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent unloading technology, specifically to a silo discharge device and its anti-blocking method based on monitoring and guidance. Background Technology

[0002] Silos are warehouses for storing bulk materials, and are divided into two main categories: agricultural silos and industrial silos. Agricultural silos are used to store granular and powdery materials such as grains and feed. Industrial silos are used to store bulk materials such as coke, cement, salt, and sugar. To ensure the efficiency of loading and unloading materials, preventing blockage at the silo inlet and outlet is an important task. For example, patent 202322611792.3 discloses an anti-clogging soybean meal silo, which is equipped with a silo body, a discharge hopper, a discharge pipe, a discharge auger, and arch-breaking rods. The drive motor of the auger discharge can drive several arch-breaking rods to break up and disperse the soybean meal material in the silo, making the soybean meal material more loose before it enters the discharge hopper for unloading, which can effectively prevent blockage during the unloading process; however, the configuration cost of the anti-clogging device of this structure is relatively high, and the modification of traditional silos is relatively complicated. Meanwhile, traditional silo inlets and outlets lack flow monitoring and cannot predict blockages. When blockages occur, they cannot be automatically and quickly resolved. They often only react when the blockage is large-scale and severe. This slow and delayed processing increases the difficulty of blockage resolution and reduces the efficiency of blockage clearing. Therefore, this invention proposes a silo discharge device and a monitoring-based anti-blockage method to solve the above problems. Summary of the Invention

[0003] The purpose of this invention is to overcome the deficiencies of the prior art and provide a silo discharge device and a monitoring-based anti-blocking method to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows: A silo discharge device includes a discharge cylinder, a guiding actuator, a ring sensor matrix, and a control center; the discharge cylinder is connected to and communicates with the silo discharge port to form a discharge channel, and the guiding actuator is installed in the discharge channel; A ring-shaped sensor matrix is ​​fixed inside the discharge channel and located above the guiding actuator. The ring-shaped sensor matrix includes N monitoring units evenly distributed circumferentially. Each monitoring unit includes a vibrating plate, a support spring, and a pressure sensor. The vibrating plate is set at an angle to the material flow direction, and its upper end is mounted with a pin and elastically supported by the support spring. The pressure sensor is arranged radially and is configured to detect pressure signals by contacting the vibrating plate. The pressure signals are processed by frequency domain separation to decouple and obtain multi-dimensional monitoring data. The multi-dimensional monitoring data includes the vibration frequency of the vibrating plate, as well as the static pressure and dynamic friction force of the material on the vibrating plate. The control center is connected to the ring sensor matrix and the dredging actuator respectively. It is configured to receive and process the pressure signals of the ring sensor matrix, and identify the flow state of the material based on the spatial characteristics, temporal characteristics and frequency domain characteristics of the multi-dimensional monitoring data of N monitoring units. The flow state includes normal, slight adhesion, high risk of bridging and blockage. The control center controls the dredging actuator to perform the corresponding dredging action according to the identified flow state.

[0005] According to one aspect of this disclosure, the diversion actuator includes a variable frequency silo vibrator and a diversion plate assembly driven by a servo motor. Depending on the flow state, the control center is configured to execute the following graded diversion strategies: When the flow status is "normal", the servo motor is controlled to run at low speed and the bin wall vibrator is turned off; When the flow state is "slight adhesion", control the servo motor to run at medium speed and start the bin wall vibrator to run at low frequency; When the flow status is "high risk of bridging", the servo motor is controlled to run at high speed and the bin wall vibrator is controlled to run at medium frequency, while triggering an early warning. When the flow status is "blocked", the servo motor is controlled to run at ultra-high speed and the bin wall vibrator is controlled to run at high frequency, while triggering an alarm.

[0006] A silo discharge device based on a monitoring and guidance method for preventing blockages includes the following steps and contents: S1, N monitoring units collect pressure signals of the flowing material in real time; S2. The control center performs frequency domain separation and decoupling on the pressure signal, generates multi-dimensional monitoring data, performs preprocessing, and finally outputs a standardized time-series data sequence. S3. Perform multi-dimensional feature extraction, fusion, reasoning, and decision-making on multi-dimensional monitoring data, among which... S31. Multidimensional features include spatial features, time-domain features, and frequency-domain features; Extracting spatial features: Based on multidimensional monitoring data from N monitoring units at the same time, calculate the uniformity of pressure, friction and vibration frequency distribution on the cross-section of the discharge channel, including pressure distribution variance and maximum pressure gradient; Extracting time-domain features: Set the sliding window size and step size, and based on the time-series data within the sliding window, calculate at least the mean static pressure, the variance of dynamic friction, the peak vibration frequency, and the rate of pressure change. Extracting frequency domain features: Performing spectral analysis on the time-series data of vibration frequencies to extract at least the energy proportion of a preset frequency band; S32. Introduce a feature-level attention mechanism to automatically assign weights to each feature dimension contained in spatial features, temporal features, and frequency features, highlighting the core features related to the flow states of "slight adhesion", "high risk of bridging" and "blockage", and output a weighted fusion feature vector. S33. Input the fused feature vector into the pre-trained blockage prediction model, perform state reasoning, and output the material state analysis results, which include the flow state, blockage trend probability, and corresponding confidence level. S4. Set the probability threshold and confidence threshold for the blockage trend to control the dredging execution mechanism to perform the corresponding graded dredging actions.

[0007] According to one aspect of this disclosure, in the initial learning phase, an update frequency is set, dynamic benchmarks for each specific feature under "normal" flow conditions are dynamically statistically analyzed, and feature thresholds used for flow condition determination are dynamically calculated and updated.

[0008] According to one aspect of this disclosure, an incremental update threshold is set, and when the number of high-quality decision samples reaches the incremental update threshold, the congestion prediction model is fine-tuned and incrementally updated using an online gradient descent algorithm.

[0009] According to one aspect of this disclosure, physical property parameters of different materials and their corresponding optimal model parameter sets are stored. For new materials, based on their physical property similarity with existing materials, the existing model parameters with the highest similarity are quickly called and migrated as initial parameters. In the initial learning stage, based on the specific feature data of the current conveyed material in the "normal" flow state, the blockage prediction model is triggered to be finely adjusted and updated incrementally, while the feature threshold for determining the flow state is dynamically updated.

[0010] According to one aspect of this disclosure, the blockage prediction model adopts a hybrid "LSTM+RF" model, which is trained using historical data. The historical data covers a variety of typical material categories, common environmental conditions and operating conditions, and is constructed into several training datasets including normal unloading scenarios, slight adhesion scenarios, bridging scenarios and blockage scenarios. The hyperparameters of the blockage prediction model are optimized using the 5-fold cross-validation method, and the model accuracy threshold and prediction lead threshold are set. The blockage prediction model is iteratively optimized.

[0011] According to one aspect of this disclosure, an unloaded test is performed before commissioning to conduct unloaded learning and calibration: the device is started in a material-free state, the average output value of the pressure sensor is dynamically counted and recorded as the "zero point value", and the "zero point value" is used as the reading calibration reference value; the time series data of vibration frequency under unloaded state is analyzed as the basis for preprocessing.

[0012] According to one aspect of this disclosure, initialization and debugging are performed synchronously before commissioning: A1. Load the pre-trained model parameters and historical feature thresholds of the currently transported material or similar material to achieve a hot start of the model; A2. Conduct small-batch material flow trial transport tests. When the flow state is determined to be "normal", collect several high-quality normal flow samples and dynamically calculate and update the feature threshold. A3. Enhance performance by artificially creating training scenarios: Artificially simulate and induce flow signs such as "slight adhesion" and "high risk of bridging", and manually label the time series data stream in real time during the simulation process, collect a number of simulated scenario data samples, and trigger incremental updates to the congestion prediction model.

[0013] Compared with the prior art, the silo discharge device and its anti-clogging method based on monitoring and guidance of the present invention have the following beneficial effects: 1. This discharge device constructs a multi-dimensional monitoring force system within the discharge channel. Using this force system as the monitoring unit of a ring sensor matrix, it can capture the non-uniformity of material force on the cross-section of the discharge channel and the temporal evolution trend of single points. The control center performs AI analysis and identification based on the monitoring data of the ring sensor matrix, outputs the flow status of the material, and controls the dredging actuator to execute a graded dredging strategy. By monitoring data in real time, it can detect early signs of blockage formation, predict the risk of blockage in advance, and take timely dredging strategies to reduce the probability of severe blockage. This strategy achieves a smooth transition and automatic degradation from undisturbed operation to powerful blockage breaking, applying minimal effective intervention only when necessary. While ensuring the dredging effect, it significantly reduces energy consumption, equipment wear, and material breakage rate. In addition, the device has a compact structure and is easy to modify. Compared with the traditional solution that requires a large number of arch-breaking devices to be installed in the silo, it has lower cost and wider applicability. 2. A method for preventing blockage based on monitoring and guidance of the discharge device: Through real-time perception and quantitative analysis of the stress and flow spatial distribution of the entire cross-section of the discharge channel and the time series of single-point data, spatial features, time series features and frequency domain features are extracted. An attention mechanism is introduced to dynamically weight different features and fuse them, so that the model can focus on the key signal changes that are strongly correlated with blockage. This improves the recognition accuracy and confidence of the flow state under different materials and different working conditions. The system can identify clear signals of bridging risk and adhesion before the material flow has completely stopped, realize early warning, and create a key time window for active intervention. 3. This method integrates perception, computation, AI reasoning, and intelligent execution to form an intelligent control closed loop, which can replace manual experience and intermittent operation, and achieve continuous, autonomous, and intelligent anti-blocking protection in the material discharge process. Attached Figure Description

[0014] Figure 1 This is a three-dimensional structural diagram of the discharge device disclosed in this invention; Figure 2 This is a cross-sectional view of the material discharge device disclosed in this invention; Figure 3 for Figure 2 A magnified view of a section at point I; Figure 4 The present invention relates to a discharge device based on a monitoring and guidance method for preventing blockage.

[0015] Attached image labels: 100. Discharge cylinder; 110. Sealing flange; 200. Guide plate assembly; 210. Servo motor; 220. Transmission bevel gear pair; 230. Guide plate; 240. Support frame; 250. Rotating rod; 300. Monitoring unit; 310. Annular base; 311. Mounting groove; 320. Static cone hopper; 330. Vibrating plate; 340. Pressure sensor; 350. Support spring. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely the best embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] The term "embodiment" as used herein means that a particular method, step, or content described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0018] This embodiment provides a silo discharge device, such as... Figures 1-3 As shown, the discharge device is coaxially connected to the silo inlet, and its upper sealing flange 110 is inserted and sealed to the silo inlet, connecting to the silo inlet. The discharge device includes a discharge cylinder 100, a guiding actuator, a ring sensor matrix, and a control center. The discharge cylinder 100 communicates with the silo inlet to form a discharge channel, the last section of which is a conical structure. The guiding actuator is located directly below the ring sensor matrix, and both are installed within the discharge channel. The dredging actuator includes a frequency-converting bin vibrator and a dredging plate assembly 200. The dredging plate assembly 200 includes a support frame 240, a servo motor 210, a transmission bevel gear pair 220, a rotating rod 250, and dredging plates 230. The support frame 240 is welded and fixed to the wall of the discharge cylinder 100 and is used to position and fix the upper end of the rotating rod 250. The servo motor 210 is located outside the discharge cylinder 100, and its drive rod extends into the discharge channel. It drives and connects to the rotating rod 250 through the transmission bevel gear pair 220. The rotating rod 250 is located on the central axis of the discharge cylinder 100, passes through the conical structure section, and extends out of the lower end of the discharge cylinder 100. Several dredging plates 230 are welded at equal intervals in the lower part of the rotating rod 250. The dredging plates 230 have a helical angle consistent with the dredging rotation direction. The frequency-converting bin vibrator is connected to the discharge cylinder 100 and applies vibration to the discharge cylinder 100 to flexibly break the arches of the material inside the cylinder. The annular sensor matrix includes N monitoring units 300 evenly distributed circumferentially. These units are mounted within the discharge channel via annular bases 310, which are coaxially welded to the inner arm of the discharge channel. N mounting slots 311 are formed on the inner annular surface of the base for mounting the monitoring units 300. Each monitoring unit 300 includes a vibrating plate 330, a support spring 350, and a pressure sensor 340. The upper end of the vibrating plate 330 is pinned to the side wall of the mounting slot 311. Its front and back surfaces are elastically supported by the support spring 350, forming a detection angle with the vertical material conveying direction. Its inclined front surface contacts the flowing material, while its back surface collects data through a detection combination coupled with the support spring 350 and the pressure sensor 340. Both the support spring 350 and the pressure sensor 340... Fixedly installed in the mounting groove 311, the pressure sensor 340 is arranged radially between the support spring 350 and the pin. The vibrating plate 330 is shielded on the opening of the mounting groove 311. It vibrates under the action of the flowing material in the discharge channel and transmits the force to the pressure sensor 340 on its back. This allows the pressure sensor 340 to detect the static pressure and dynamic friction of the vibrating plate 330 and obtain the vibration frequency based on the vibration signal of the vibrating plate 330. Based on the phenomenon that the physical process of "bridging" leaves a recognizable unique data feature pattern on the circumferentially distributed matrix of pressure sensors 340, adjacent vibrating plates 330 are flexibly connected to form a multi-point independent data acquisition unit that is evenly distributed in the circumference, so as to accurately collect matrix data for predicting the bridging phenomenon. like Figure 2 and Figure 3 As shown, a static cone hopper 320 is also configured on the discharge channel, which is coaxially set directly below the annular base 310 to guide the discharge of controllable flowing materials. The control center is connected to the ring sensor matrix and the dredging actuator respectively. It is configured to receive and process multi-dimensional time-series data collected by the ring sensor matrix from N spatial locations. Specifically, it captures the non-uniformity of material force on the cross-section of the discharge channel by calculating the spatial statistics (such as pressure distribution variance) of the data from N monitoring units 300 at the same time. Combined with the time-series evolution trend of the data at each point, it comprehensively identifies the flow state of the material. The flow state includes normal, slight adhesion, high risk of bridging and blockage. Based on the flow state identification result, it controls the dredging actuator to perform the corresponding dredging action. The control center includes an edge computing gateway, an AI analysis unit, and a monitoring display screen. The edge computing gateway collects data from the ring sensor matrix, performs filtering, fusion, and preliminary feature extraction preprocessing, and then sends the data to the AI ​​analysis unit in the cloud via wired or wireless means for flow status diagnosis. The AI ​​analysis unit is equipped with a trained blockage prediction model to determine the material flow status as "normal", "slight adhesion", "high risk of bridging" and "blocked". The real-time status information is displayed on the monitoring display screen. At the same time, based on the flow status identification, the control center sends corresponding control signals to the dredging actuator to execute the graded dredging strategy.

[0019] This embodiment also provides a method for preventing blockage of a silo based on the above-described discharge device configuration, such as... Figure 4 As shown, the flow status is identified and diversion is performed based on monitoring data from a ring-shaped sensor matrix. Specifically, it includes the following steps and contents: S1. The ring sensor matrix serves as a sensing module. It consists of a vibrating plate 330, a support spring 350, and a pressure sensor 340, forming a multi-dimensional force monitoring system. When the material flows through it, it squeezes the vibrating plate 330 that extends obliquely from the inner ring surface. The output of the pressure sensor 340 connected to the vibrating plate 330 couples the pressure signals of the three functions. S2. The edge computing gateway performs real-time frequency domain separation and decoupling processing on the pressure signal to extract three independent physical quantities from the coupled original signal: the static pressure P of the flowing material on the vibrating plate 330, the dynamic friction force F, and the vibration frequency f of the vibrating plate 330 excited by the material flow. The specific decoupling process is carried out sequentially according to the physical meaning of the signal component separation: First, extract the static pressure P: Static pressure P reflects the degree of material accumulation at the monitoring point, corresponding to the near-DC component in the signal. The raw pressure signal is passed through a low-pass filter with an extremely low cutoff frequency (e.g., 0.1-0.5 Hz) to filter out all AC fluctuations. The resulting smooth signal baseline is the time-series value of the static pressure P.

[0020] Secondly, the vibration frequency f is extracted: the vibration frequency f reflects the agglomeration state and flow energy level of the material, and its energy is concentrated in a specific frequency band related to the material characteristics. A bandpass filter preset according to the current material characteristics (for example, for particulate materials, the passband may be 5-50 Hz) is applied to filter the original signal, retaining the principal vibration components. Subsequently, a fast Fourier transform is performed on the filtered signal to identify the main peak frequency in its amplitude spectrum, which is the instantaneous value of the vibration frequency f.

[0021] Finally, the dynamic friction force F is extracted. The dynamic friction force F reflects the resistance and viscosity characteristics of material flow, manifesting as a broadband random fluctuation superimposed on static pressure and periodic vibration. By simultaneously subtracting the time-series value of the static pressure P obtained in the previous steps and the vibration signal component extracted through bandpass filtering from the original pressure signal, a residual AC signal is obtained. This residual signal undergoes further band-limited filtering to remove high-frequency noise, and its amplitude envelope is extracted using an envelope detection algorithm. Finally, based on the force-electric calibration coefficients of the sensor system, the amplitude of this envelope is converted into the time-series value of the dynamic friction force F.

[0022] In the early stages of blockage formation, the amplitudes of the above three types of data and their spatial distribution on the annular sensor matrix often exhibit abnormal gradients, such as sudden drops in local pressure (indicating cavitation), general increases in friction (decreased fluidity), decreases in vibration frequency (weakened kinetic energy), or uneven distribution of matrix data (flow instability). By analyzing the temporal trends and spatial distribution evolution of these multidimensional data, early prediction of blockages can be achieved.

[0023] Subsequently, the decoupled multidimensional monitoring data sequence was processed. (Where i is the number of the i-th pressure sensor 340) Preprocessing is performed: First, environmental noise (such as overall silo vibration and electrical interference) is suppressed using the Kalman filter algorithm; then, abnormal data points caused by instantaneous sensor failures are removed according to the 3σ criterion; for missing values ​​caused by removal or communication loss, linear interpolation is used to complete them. Finally, a cleaned and normalized standardized time-series data sequence is output. To reduce the computational load on the cloud-based AI analysis unit, the edge computing gateway performs preliminary statistical feature extraction after data preprocessing. For example, it calculates the mean pressure μP and friction variance σF for each monitoring channel within a sliding window. 2 and peak vibration frequency f max These statistical features can serve as supplementary inputs for subsequent advanced feature analysis; S3. In order to achieve early congestion warning by utilizing the spatially distributed measurement information provided by the ring sensor matrix, the AI ​​analysis unit performs the following steps: S31. Based on the above multidimensional monitoring data, multidimensional feature extraction is performed, extracting spatial features, time-domain features and frequency-domain features respectively; Spatial Feature Extraction: Spatial features are used to capture cross-sectional flow field anomalies. Based on the unique spatial deployment of the ring sensor matrix, it collects monitoring data from N monitoring units 300 at the same time, and calculates the spatial distribution uniformity of pressure, friction, or vibration frequency along the circumference of the discharge channel, including pressure distribution variance. The intelligent identification of "high bridging risk" flow states, based on the maximum pressure gradient, is mainly based on structural anomalies appearing in the spatial distribution characteristics, including the pressure distribution variance. As a core indicator for quantifying pressure voids, the variance of pressure distribution is measured when local voids occur. The pressure gradient will increase rapidly and significantly; the maximum pressure gradient is the key indicator for detecting pressure ridges. Calculate the maximum absolute value of the pressure difference between 300 adjacent monitoring units of the ring sensor matrix. When a sharp pressure ridge appears, the maximum pressure gradient will suddenly increase to an abnormally high level. At the same time, the pressure distribution curve will show clear peaks (i.e., high-pressure ridges) and troughs (i.e., low-pressure cavities). Time-domain feature extraction: Time-domain features characterize instantaneous and short-term dynamics. For each pressure sensor's 340 data stream, a 5-second sliding window with a 1-second step size is used to extract the time-series data within each sliding window, obtaining trend characteristics (such as pressure change rate). Frictional tendency Rate of change of frequency The system extracts waveform features (root mean square of pressure signal, peak factor of friction signal, etc.) and uses the initially extracted statistical features as a supplement to finally obtain 12-dimensional time-domain features. Frequency domain feature extraction: Frequency domain features reveal the intrinsic state of materials, based on the vibration frequency time series data of each pressure sensor 340. Within the same sliding window, a fast Fourier transform is performed to extract the energy proportion features of a preset low-frequency band (0-5Hz). In addition, the main frequency position and spectral entropy can also be extracted. S32. Introducing a feature-level attention mechanism to achieve adaptive weighted fusion of features. Before fusion, the spatial feature vector, temporal feature vector, and frequency domain feature vector extracted at time t need to be standardized to eliminate the training bias of the neural network model caused by differences in the dimensions, orders of magnitude, and distribution of different features. The standardized three types of feature vectors are concatenated to form a total feature vector, where each dimension corresponds to a specific feature. The total feature vector is input into a learnable linear projection layer to generate a query vector, key vector, and value vector. The similarity between the query vector and all key vectors is calculated to obtain an attention score matrix. This matrix is ​​then normalized using the softmax function and transformed into an attention weight matrix. The attention weight matrix is ​​used to sum all value vectors in a weighted manner to output a 32-dimensional fused feature vector that fuses the global feature dependencies. For example: When the material flow is in a "high bridging risk" state, the attention weight matrix displays the pressure change rate. and pressure distribution variance The weights of parameters such as pressure change rate increase significantly, which is similar to the characteristic trend of bridge-building dynamics. The values ​​are significantly negative and the variance of the pressure distribution is high. As the pressure begins to increase, the attention mechanism automatically assigns a rate of change to the stress level. and pressure distribution variance Higher weighting is given to low-frequency energy characteristics, which are also given higher weighting to confirm changes in material state. When the material flow is in a "slightly sticky" state, the attention weight matrix shows the friction variance. The weighting of factors such as the proportion of low-frequency energy has increased significantly. Based on the physical process of adhesion, the local fluidity of the material decreases, and the overall flow may slightly decrease, but no structural flow obstruction has yet formed, i.e., the average pressure of the local pressure sensor 340. There may be a slight increase, while the variance of friction... It will significantly increase the peak vibration frequency. There was a slight decrease, but the proportion of low-frequency energy showed an upward trend, indicating that the attention mechanism learned to work within the friction variance. When both the low-frequency energy proportion and the low-frequency energy proportion exhibit abnormal changes, significantly increase the weights of both, and adjust the friction variance. The characteristic of low-frequency energy proportion receives the highest weight, while the characteristic of low-frequency energy proportion receives the second highest weight. When material flow is in a "blocked" state, the flow has completely stopped, bridging or compaction blockage has formed, and the pressure distribution variance... Reaching a maximum value, the maximum pressure gradient is at an extremely high value, and both are dominant features. The attention mechanism increases the weights of both, and the rate of change of pressure. As an auxiliary judgment, it receives the second highest weight, and at the same time, the peak value of the vibration frequency is automatically increased. Weighting, upon confirming that the flow kinetic energy is low, automatically reduces the frictional variance. Weight; For a "normal" flow state, the attention mechanism distributes the weights of various features equally. S33. Construct a blockage prediction model based on the "LSTM+RF" cascaded hybrid model architecture. The training data comes from historical datasets collected by experimental devices or field equipment equipped with a ring sensor matrix. This dataset covers "multiple material types, multiple operating conditions, and multiple states." The materials should include powdery materials (such as cement, coal powder, and flour), granular materials (such as wheat and corn), and lumpy materials (such as coke and plastic granules), with at least five typical bulk materials that may be applicable. Then, for each material, the model is tested under different ambient humidity (8%-25%) and different material levels. Data was collected at a temperature of 30%-100% of the silo height to simulate fluctuations in real production. For the collected data samples, combined with the multi-dimensional monitoring data of the pressure sensor 340 and the actual unloading results, the material flow status was manually labeled into four categories: normal, slight adhesion, high risk of bridging, and blocked. Training datasets were constructed that included several groups of normal unloading scenarios, slight adhesion scenarios, bridging scenarios, and blocked scenarios under multiple materials, multiple working conditions (different unloading speeds, different material storage heights, etc.), and multiple environments (dry or different humidity levels). Using at least 2,000 sets of independent cross-scenario historical data, including at least 300 sets of scenarios with slight adhesion, high bridging risk, and blockage, the scenario data, status labels, raw time-series data collected by pressure sensor 340, and 32-dimensional fused feature vectors are input into the blockage prediction model for training. Five-fold cross-validation is used and the model and hyperparameters are iteratively optimized based on evaluation metrics. The hybrid architecture congestion prediction model and attention mechanism are jointly trained end-to-end using training set data. The model performance is evaluated using validation set. Evaluation metrics include, but are not limited to, model accuracy (e.g., 95% or higher) and prediction lead (not less than 3 seconds). The optimal hyperparameters are selected and the final congestion prediction model for deployment is determined. The LSTM layer receives a 32-dimensional fused feature vector, capturing the progressive evolution of dangerous temporal patterns of blockage risk. It learns to identify "normal," "slight adhesion," "high bridging risk," and "blocked" modes, encoding information on the coordinated deterioration trends of spatial, temporal, and frequency domain features, and outputting a trend context vector. The RF layer concatenates the trend context vector with the fused feature vector at the current moment to form a context-enhanced final feature vector. The corresponding decision trees in the RF layer are activated simultaneously, and the decision rules during decision-making are based on the multi-feature combination patterns of the four scenarios learned during training. Finally, a structured material state analysis result is output, including flow state, confidence level C, and blockage trend probability P. d (i.e., the probability of blockage) For flow state labels such as "slight adhesion" and "high risk of bridging", output a probability value of blockage occurring within the next T seconds (T is a configuration setting, and data is extracted and labeled based on this configuration setting during training). This probability value is calculated by the fully connected output layer of the model. S4. Based on the above material state analysis results, intelligent decisions are output based on dual thresholds, and a graded guidance strategy is implemented to smoothly transition from undisturbed operation to strong intervention: The flow status is "normal", with a confidence level C ≥ 0.9, and the probability of blockage trend P. d For any value, control the servo motor 210 to run at low speed and turn off the bin vibrator; The flow state is "slight adhesion", with a confidence level C ≥ 0.85 and a blockage trend probability P. d <60%, the confidence level meets the standard but the risk does not reach the threshold, so control the servo motor 210 to run at medium speed and start the bin vibrator to run at low frequency; The flow status is "high risk of bridging", with confidence level C ≥ 0.85 and blockage trend probability P. d If the probability of failure is ≥60%, the probability warning condition is met. The servo motor 210 is controlled to run at high speed and the bin wall vibrator is controlled to run at medium frequency, and a warning is issued. The flow status is "blocked" with a confidence level C≥0.85. At this point, the blockage is a fait accompli. The servo motor 210 is controlled to run at ultra-high speed and the bin wall vibrator is controlled to run at high frequency, while triggering an alarm. It should be understood that the above-mentioned "low speed", "medium speed", "high speed", "ultra-high speed" and "low frequency", "medium frequency" and "high frequency" are relatively divided or specified according to the speed range or speed value preset by the servo motor 210 and the frequency range or frequency value preset by the bin vibrator. Compared to existing technologies that rely on single-point monitoring or time-series analysis, this embodiment creatively utilizes a ring-shaped sensor matrix to achieve real-time sensing and quantitative analysis of the stress and flow spatial distribution across the entire cross-section of the discharge channel, and extracts the pressure distribution variance. Based on spatial characteristics such as maximum pressure gradient, the system can detect abnormal mechanical distribution of the marker bridging structure before the material flow has completely stopped. As a result, the control center has the ability to make early diagnoses based on spatial distribution patterns. Combined with time-series trend analysis and multi-modal fusion prediction, it realizes intelligent anti-blocking material discharge control with high precision, early warning of blockage, and early intervention.

[0024] As a further technical solution, to overcome the problem of fixed threshold failure caused by material property differences, environmental changes, and equipment state drift, the AI ​​analysis unit is also equipped with a dynamic threshold adjustment module. This module dynamically calculates and updates the relative characteristic thresholds used for state determination based on the "normal" flow state reflected by the most recent historical data. Specifically: after the system starts unloading or material switching, it enters an initial learning phase (e.g., the first 5 minutes of stable operation), continuously collecting characteristic data. When the blockage prediction model determines "normal" with high confidence, the characteristic data at that moment is included in the baseline data buffer. An update frequency is set, and the dynamic baselines for each characteristic within the baseline data buffer are dynamically calculated, generating relative thresholds in real time for judging each flow state. For example, the average pressure within the baseline data buffer is calculated. The average value of all sliding windows within a 1-minute update frequency is used as the baseline average. The baseline fluctuation range is then calculated based on this baseline average. and benchmark fluctuation range Generate judgment threshold For example, in the "slight adhesion" criterion, k is taken as 1.5-2, while in the "high bridge construction risk" criterion, k is a more aggressive coefficient, and the average pressure... When the flow rate falls below any threshold, it indicates that the material is in the early stages of flow blockage and its flowability is beginning to decline. Based on this, a collaborative system formed by dynamic threshold updates based on multiple features can be used to predict the flow state in an adaptable manner to changes in material properties, environment, and equipment status.

[0025] Furthermore, as a further technical solution, to achieve continuous performance evolution throughout the entire lifecycle, the AI ​​analysis unit is also equipped with an adaptive learning module. This module enables incremental online learning of model parameters and the construction and migration of material property databases. During system operation, high-quality decision samples are continuously collected and stored in the incremental learning sample set. High-quality decision samples refer to samples with high confidence that have been manually confirmed or verified by feedback. When the number of samples in the incremental learning sample set reaches the incremental update threshold (e.g., 50 sets), the incremental learning online learning mechanism is triggered. The online gradient descent algorithm is used to jointly optimize the prediction model and attention mechanism with 50 sets of new data as the incremental training set. Forward and backward propagation are performed with a small learning rate to update the weight parameters of the LSTM layer, dynamically update the tree node statistics, or add new decision trees to update the rules of the existing forest. The system stores the physical property parameters of different materials and their corresponding optimal model parameter sets. For new materials, based on their physical property similarity with existing materials, it quickly calls and transfers the existing model parameters with the highest similarity as initial parameters. During the initial learning phase, based on the specific feature data of the current material under "normal" flow conditions, it triggers fine-tuning and incremental updates to the blockage prediction model, while dynamically updating the feature thresholds for flow state determination. A material property database is established for each successfully operated material, linking and storing the material's physical property parameters (such as angle of repose, density, particle size distribution, etc.) with its optimal model parameter set. When a new material is introduced for the first time, its physical property similarity with existing materials in the material property database is calculated by measuring the physical property parameters of the measuring instrument. For example, the Euclidean distance of key physical property parameters is calculated. Based on the physical property similarity, existing model parameters are quickly called and selected as initial parameters. For example, for grain materials, the vibration frequency characteristics are emphasized, and for powdery materials, the friction characteristics are emphasized. The unloading system is started to discharge the material. During the trial operation of the new material (such as within 10-30 minutes of material conveying), an incremental online learning mechanism is triggered. The small amount of data generated during the trial operation is used to quickly fine-tune the hot-start model until it converges to the optimal state for the new material.

[0026] As a further technical solution, before the device and system are put into use, pre-start checks, no-load learning and calibration, and system initialization are required. (1) Pre-start checks: Mechanical component inspection: Manually rotate the drive rod of the servo motor 210 and observe whether the meshing of the transmission bevel gear to 220 is smooth, whether the guide plate 230 is not stuck, and check whether the connection between the discharge cylinder 100 and the silo opening is firm and sealed. Electrical system inspection: Check whether the power supply of each pressure sensor 340, servo motor 210, and frequency converter is normal; test whether the pressure sensor 340 can collect data normally; test whether the AI ​​analysis unit can receive and process data normally; and test whether it can send control signals normally. (2) Idle learning and calibration: Initiate an unloaded test with no material present. Calculate the average output value of pressure sensor 340 using the dynamic threshold adjustment module and record it as the "zero point value." Use this zero point value to calibrate the reading of pressure sensor 340. Analyze the vibration frequency data under unloaded conditions and record the spectrum data as a background noise template in preprocessing. Record the fluctuation range of the data collected by pressure sensor 340 under unloaded conditions as the "static noise threshold" for signal filtering in preprocessing. (3) System initialization: The AI ​​analysis unit is put into use only after structured initialization and debugging, and the deep binding and calibration of the algorithm with the device, environment and materials are completed; A1. Import the pre-trained model parameters and historical feature thresholds of the target material or similar physical properties from the cloud or local database, load them into the AI ​​analysis unit, and complete the model hot start. A2. Conduct small-batch material flow trial transport tests. The guiding components run at low speed and continuously collect data. When the AI ​​analysis unit continuously and stably determines the state as "normal", these data are considered as high-quality normal samples. At least 100 sets of high-quality normal flow samples are collected, with a 5-second sliding window of data as one sample. The judgment threshold of specific features is recalculated and updated using the dynamic threshold adjustment module. A3. Performance Enhancement: Artificially create training scenarios to achieve working condition simulation and active learning. By artificially adjusting the simulation to induce flow signs such as "slight adhesion" and "high risk of bridging", such as by briefly stopping the material and spraying water to increase humidity, the data stream is manually labeled in real time during the simulation. For example, when the material begins to show viscous and slow flow, click "slight adhesion"; when there is obvious flow obstruction, click "high risk of bridging". Collect a certain number of scenario data sets (30 sets of labeled samples covering "normal", "slight adhesion" and "high risk of bridging") to trigger the above self-learning module to fine-tune the model parameters.

[0027] Application Example 1: Application Background: A large grain storage facility uses concrete silos with a diameter of 5m and a height of 20m to store wheat. The bottom opening of the silo has a diameter of 800mm. During the traditional discharge process, bridging and blockage often occur due to changes in the moisture content of the wheat of 12%-15%, resulting in interruption of discharge and requiring manual cleaning, which is inefficient. Device adaptation and installation: Based on the material inlet size, the top inner diameter of the discharge cylinder 100 is customized to be 800mm, the bottom inner diameter to be 600mm, the total length to be 1.2m, and the lower section length to be 0.8m; the annular base 310 is made of stainless steel and is adapted to the inner diameter of the discharge cylinder 100; the pressure sensor 340 is a high-precision sensor with a range of 0-50kPa, and a total of 12 sensors are set to form an annular sensor matrix; the servo motor 210 is a variable frequency motor with a power of 3kW and a rated speed of 1500r / min; the guide plate 230 is made of wear-resistant manganese steel, and 4 plates are set, each plate with a width of 150mm, to ensure that the cross-section of the discharge cylinder 100 is covered; Application Results: The discharge device can monitor pressure changes during the wheat flow process in real time. When the wheat slightly sticks together, the AI ​​analysis unit detects the friction variance for the first time. An abnormal rise in temperature is identified as "slight adhesion" within 3 seconds, triggering a low-level drainage system. When moderate blockage occurs, the algorithm determines a "high risk of bridging" with a confidence level of 0.92 and a blockage trend probability of 80%. The servo motor speed is automatically increased to 180 r / min, and the vibrator frequency is set to 35 Hz. Normal discharge can be restored in an average of 15-20 seconds. Compared to the traditional manual cleaning method, which takes an average of 30-60 minutes, the drainage efficiency is improved by more than 90%. In addition, the "flexible" arch-breaking method using a variable frequency silo wall vibrator controls the breakage rate of wheat during the drainage process to below 0.5%, which is far lower than the 2%-3% breakage rate of traditional mechanical arch-breaking.

[0028] Application Example 2: Application Background: A cement production plant uses steel silos with a diameter of 8m and a height of 30m to store cement. The feed pipe at the bottom of the silo has a diameter of 600mm. Because cement is prone to absorbing moisture and clumping, blockages often occur during the feeding process, causing feeding interruptions and affecting the continuity of production. Device adaptation and installation: The customized discharge cylinder 100 has a top inner diameter of 600mm, a bottom inner diameter of 500mm, a total length of 1.0m, and a conical section length of 0.6m. Considering the corrosiveness of cement, the annular base 310 and the guide plate 230 are made of 316 stainless steel. The pressure sensor 340 is a dustproof and corrosion-resistant encapsulated model, with a total of 8 sensors forming an annular sensor matrix. The servo motor 210 is a 2.2kW variable frequency motor with a braking function to prevent the guide plate 230 from reversing due to the gravity of cement when the machine stops. Wear-resistant liners are added to the inner wall of the discharge cylinder 100 to extend its service life. Application results: During three consecutive months of operation, the blockage rate was reduced from the traditional 8-10 times per month to 1-2 times. When a blockage occurs, the servo motor 210 drives the guide plate 230 to rotate at a speed of 250 r / min. Combined with the frequency conversion bin vibrator with a vibration frequency of 50 Hz, the blockage channel can be cleared in an average of 8-12 seconds, ensuring continuous feeding and reducing production interruption losses. Cross-scenario adaptation: When cement absorbs moisture and clumps, the AI ​​analysis unit automatically triggers the adjustment of feature weights (friction feature weight increases from 0.25 to 0.38) after accumulating 50 sets of data through the adaptive learning module, adapting to changes in material state without manual intervention; Blockage prediction: The AI ​​analysis unit detected abnormal changes in the pressure change rate and the proportion of low-frequency vibration energy during the cement feeding process, determined it to be "high risk of bridging" and initiated intermediate-level drainage. Compared with human intervention, it predicted the risk of bridging 7 minutes in advance. Algorithm stability: During three consecutive months of operation, the model optimized parameters 12 times through adaptive updates, and the overall monitoring accuracy remained above 98.5%, with no performance degradation caused by changes in material characteristics or environment. Maintenance cost: The discharge device is made of wear-resistant and corrosion-resistant materials, and only requires simple cleaning and maintenance once a month, which takes about 30 minutes. Compared with traditional manual cleaning, which takes 2-3 hours each time, the maintenance cost is reduced by more than 60%.

[0029] Through the above description of the embodiments, those skilled in the art can clearly understand that the various embodiments of this application can be implemented by means of software or software combined with necessary general-purpose hardware platforms, and of course, they can also be implemented by hardware functions. Based on this understanding, the technical solution of this application, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. The software product is stored in a storage medium and includes several instructions to cause a computer device, such as including but not limited to a personal computer, server, or network device, to execute all or part of the steps of the method described in any embodiment of this application.

[0030] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A silo discharge device, characterized in that: It includes a discharge cylinder, a guiding actuator, a ring sensor matrix, and a control center; the discharge cylinder is connected to and communicates with the silo inlet to form a discharge channel, and the guiding actuator is installed in the discharge channel; The annular sensor matrix is ​​fixed within the discharge channel and located above the guiding actuator. The annular sensor matrix includes N monitoring units evenly distributed circumferentially. Each monitoring unit includes a vibrating plate, a support spring, and a pressure sensor. The vibrating plate is angled to the material flow direction, with its upper end pin mounted and elastically supported by the support spring. The pressure sensor is arranged radially and configured to detect pressure signals by contacting the vibrating plate. Frequency domain separation processing is performed based on the pressure signals to decouple and obtain multi-dimensional monitoring data. The multi-dimensional monitoring data includes the vibration frequency of the vibrating plate, and the static pressure and dynamic friction force of the material on the vibrating plate. The control center is connected to the ring sensor matrix and the dredging actuator respectively. It is configured to receive and process the pressure signal from the ring sensor matrix, and identify the flow state of the material based on the spatial characteristics, temporal characteristics and frequency domain characteristics of the multi-dimensional monitoring data of the N monitoring units. The flow state includes normal, slight adhesion, high risk of bridging and blockage. The control center controls the dredging actuator to perform the corresponding dredging action according to the identified flow state.

2. The silo discharge device according to claim 1, characterized in that: The diversion actuator includes a variable frequency bin vibrator and a diversion plate assembly driven by a servo motor. Based on the flow state, the control center is configured to execute the following tiered diversion strategy: When the flow state is "normal", the servo motor is controlled to run at low speed and the bin wall vibrator is turned off; When the flow state is "slight adhesion", the servo motor is controlled to run at medium speed and the bin wall vibrator is started to run at low frequency; When the flow state is "high risk of bridging", the servo motor is controlled to run at high speed and the bin wall vibrator is controlled to run at medium frequency, while triggering an early warning. When the flow status is "blocked", the servo motor is controlled to run at ultra-high speed and the bin wall vibrator is controlled to run at high frequency, while triggering an alarm.

3. A method for preventing blockage based on monitoring and guidance using the silo discharge device of claim 1, characterized in that, Includes the following steps and content: S1, N monitoring units collect the pressure signal of the flowing material in real time; S2. The control center performs frequency domain separation and decoupling on the pressure signal, generates the multi-dimensional monitoring data, performs preprocessing, and finally outputs a standardized time-series data sequence. S3. Perform multi-dimensional feature extraction, fusion, reasoning, and decision-making on the multi-dimensional monitoring data, wherein... S31, The multidimensional features include spatial features, time-domain features, and frequency-domain features; Extract the spatial features: Based on the multidimensional monitoring data of N monitoring units at the same time, calculate the uniformity of the distribution of pressure, friction and vibration frequency on the cross-section of the discharge channel, including the pressure distribution variance and the maximum pressure gradient; Extract the time-domain features: Set the sliding window size and step size, and based on the time-series data within the sliding window, calculate at least the static pressure mean, dynamic friction variance, vibration frequency peak value, and pressure change rate; Extracting the frequency domain features: Performing spectral analysis on the time-series data of the vibration frequency to extract at least the energy proportion of a preset frequency band; S32. Introduce a feature-level attention mechanism to automatically assign weights to each feature dimension contained in the spatial features, the temporal features, and the frequency features, highlighting the core features related to the flow states of "slight adhesion", "high risk of bridging" and "blockage", and output a weighted fusion feature vector. S33. Input the fused feature vector into the pre-trained blockage prediction model, perform state reasoning, and output the material state analysis results, which include flow state, blockage trend probability and corresponding confidence level. S4. Set the blockage trend probability threshold and confidence threshold, and control the dredging execution mechanism to perform the corresponding graded dredging actions.

4. The silo discharge device according to claim 3, characterized in that: In the initial learning phase, the update frequency is set, the dynamic benchmarks of each specific feature under the "normal" flow state are dynamically statistically analyzed, and the feature thresholds used for the determination of the flow state are dynamically calculated and updated.

5. The silo discharge device according to claim 4, characterized in that: Set an incremental update threshold. When the number of high-quality decision samples reaches the incremental update threshold, fine-tune the congestion prediction model using an online gradient descent algorithm.

6. The silo discharge device according to claim 5, characterized in that: The system stores the physical property parameters of different materials and their corresponding optimal model parameter sets. For new materials, based on their physical property similarity with existing materials, it quickly calls and migrates the existing model parameters with the highest similarity as initial parameters. During the initial learning phase, based on the specific feature data of the current conveyed material in a "normal" flow state, it triggers the fine-tuning incremental update of the blockage prediction model and dynamically updates the feature threshold for the flow state determination.

7. The silo discharge device according to claim 3, characterized in that: The blockage prediction model adopts an "LSTM+RF" hybrid model, which is trained using historical data. The historical data covers a variety of typical material categories, common environmental conditions and operating conditions, and is constructed into several training datasets, including normal unloading scenarios, slight adhesion scenarios, bridging scenarios and blockage scenarios. The hyperparameters of the blockage prediction model are optimized using the 5-fold cross-validation method. The model accuracy threshold and prediction lead threshold are set, and the blockage prediction model is iteratively optimized.

8. The silo discharge device according to claim 6, characterized in that: Before commissioning, an unloaded test is performed to conduct unloaded learning and calibration: the device is started in a material-free state, and the average output value of the pressure sensor is dynamically counted and recorded as the "zero point value". The "zero point value" is used as the reading calibration reference value. The time series data of the vibration frequency under the unloaded state is analyzed as the basis for the preprocessing.

9. The silo discharge device according to claim 8, characterized in that: Initialization and debugging are performed synchronously before commissioning: A1. Load the pre-trained model parameters and historical feature thresholds of the currently transported material or similar material to achieve a hot start of the model; A2. Conduct small-batch material flow trial transport tests. When the flow state is determined to be "normal", collect several high-quality normal flow samples and dynamically calculate and update the feature threshold. A3. Artificially create training scenarios to enhance performance: Artificially simulate and induce flow signs such as "slight adhesion" and "high risk of bridging", and manually label the time series data stream in real time during the simulation process, collect a number of simulated scenario data samples, and trigger incremental updates to the blockage prediction model.

Images