Material pushing state real-time discrimination method and system based on moment sensing

Abstract

The invention relates to the technical field of material pushing mechanism control and recognition, in particular to a material pushing state real-time judgment method and system based on torque sensing and a material pushing device used for pushing products in a material box out layer by layer, and the method comprises the steps that in the material pushing process, the output torque value of a servo motor is collected in real time, and a real-time torque sequence is formed; inputting the time sequence convolutional neural network model into a pre-trained time sequence convolutional neural network model; the model extracts dynamic features of different time scales through an expansion causal convolutional layer, and outputs confidence degrees of three pushing states of empty pushing, effective load and interlayer transition; when the confidence coefficient of the empty pushing state exceeds a preset condition, it is judged that the material box is empty, and material pushing is stopped; when the inter-layer transition state confidence meets the condition, triggering an inter-layer switching operation; according to the method, the material pushing state can be accurately recognized in real time, misjudgment and delay are avoided, the material pushing precision and reliability are improved, fatigue damage of mechanical parts is reduced, and the equipment reliability and the production efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of control and identification technology for feeding mechanisms, and in particular to a method and system for real-time determination of feeding status based on torque sensing. Background Technology

[0002] Layer-by-layer pushing equipment is a type of automated device commonly used in automatic feeding, sorting, or conveying processes. Its basic function is to push products stacked vertically in a predetermined order from the material box and convey them to the subsequent workstation. Layer-by-layer pushing means that the pushing mechanism only applies pushing force to the current target layer of products in each pushing cycle, causing the products of that layer to leave the material box. After pushing is completed, a layer switching process is required to create conditions for pushing the next layer. For example, a small pushing mechanism can be equipped with a material box that can move up and down, so that the pusher head is aligned with the products in the material box at different height positions, and drives the material box to rise and fall after each pushing, so as to realize layer-by-layer pushing and material box recycling. To determine the pushing status, this type of device often relies on pushing count, pusher head stroke position, limit switch, or photoelectric detection element at the pushing outlet for status judgment.

[0003] In existing technologies, feeding equipment often uses position sensors or vision sensors to determine the feeding status. Position sensors determine the feeding progress by monitoring changes in the position of the pusher head, while vision sensors detect the quantity and position of products in the material box through image processing. Although these methods are effective in some cases, they have limitations. Position sensors cannot accurately reflect changes in torque, especially under high-frequency or complex working conditions, and are prone to errors, affecting the accurate judgment of the feeding status. Although vision sensors can provide real-time images of the product, they are greatly affected by factors such as ambient light and the surface shape of the item, and have high deployment and maintenance costs, making them difficult to adapt to various working environments. In addition, many existing systems rely on timer or switch control and cannot make dynamic adjustments according to the actual feeding status, which can easily lead to dry pushing or over-pushing, affecting feeding efficiency and equipment lifespan.

[0004] The information disclosed in this background section is intended only to enhance the understanding of the general background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0005] This invention provides a method and system for real-time determination of material pushing status based on torque sensing, thereby effectively solving the problems in the background art.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a real-time material pushing state discrimination method based on torque sensing, used in a material pushing device to push products out of a material box layer by layer, the method comprising: During the process of the pushing device performing a single-layer pushing action, the output torque sampling value of the servo motor driving the push head is collected in real time to form a real-time torque sequence; The torque sequence is input into a pre-trained push state discrimination model; the push state discrimination model is a temporal convolutional neural network, which synchronously extracts dynamic features at different time scales from the real-time torque sequence through stacked dilated causal convolutional layers, and outputs the confidence of at least three push states, including empty push state, effective load state and inter-layer transition state. Decisions are made based on the confidence level. When the confidence level of the empty push state meets the first preset condition, it is determined that the material box is empty and the push mechanism is controlled to stop the cycle. When the confidence level of the interlayer transition state meets the second preset condition, the push mechanism is triggered to perform an interlayer switching operation.

[0007] Furthermore, before inputting the real-time torque sequence into the pre-trained material pushing state discrimination model, an adaptive stroke reconstruction step is also included: The servo motor encoder signal is synchronously acquired at the moment corresponding to the output torque sampling value; Based on the real-time torque sequence, the starting time of this pushing action is determined; From the real-time torque sequence, the starting point at which the torque value first decreases from the stable load platform is detected is recorded as the torque decoupling point; The encoder position corresponding to the torque release point is taken as the start point of the stroke, and the encoder position corresponding to the torque release point is taken as the effective physical end point of the stroke. Based on the actual displacement between the starting point of the journey and the end point of the effective physical journey, normalization processing is performed to obtain a normalized journey sequence; Based on the normalized stroke sequence, the real-time torque sequence is resampled and interpolated to generate a standard stroke torque sequence of fixed length; The standard stroke torque sequence serves as the input to the material pushing state discrimination model.

[0008] Furthermore, the temporal convolutional neural network includes an input layer, a multi-path feature extraction module, and a classification output layer connected in sequence; The multi-path feature extraction module includes a transient feature extraction sub-network, a trend feature extraction sub-network, and a frequency domain feature extraction sub-network that are configured in parallel. The transient feature extraction subnetwork includes a causal convolutional layer with a first set of dilation coefficients configured to have a receptive field on a millisecond time scale, for extracting torque abrupt features characterizing the pusher contacting the product and the product leaving the hopper from the standard stroke torque sequence. The trend feature extraction subnetwork includes a causal convolutional layer with a second set of expansion coefficients, which are greater than the first set of expansion coefficients. It is configured to have a receptive field that covers the main stroke time scale of a single push, and is used to extract load plateau and trend features that characterize the stable push process of the entire layer of product from the standard stroke torque sequence. The frequency domain feature extraction subnetwork includes a time-frequency transform layer and subsequent convolutional layers, used to extract frequency domain energy distribution features characterizing the sliding friction vibration and abnormal impact of the product from the transform domain of the standard stroke torque sequence.

[0009] Furthermore, the temporal convolutional neural network also includes an attention fusion layer located between the multi-path feature extraction module and the classification output layer; the attention fusion layer is configured to perform the following operations: The feature maps output by the transient feature extraction subnetwork, the trend feature extraction subnetwork, and the frequency domain feature extraction subnetwork are concatenated along the channel dimension to obtain a concatenated feature map sequence. The normalized travel sequence is mapped to a travel position encoding vector that matches the length of the spliced ​​feature map sequence using a sine-cosine position encoding function; The concatenated feature map is added to the travel position encoding vector, and then a two-layer feedforward neural network is used to generate channel attention weights and sequence attention weights, respectively. The channel attention weight and sequence attention weight are applied to the spliced ​​feature map for weighted fusion to obtain a fused feature map, which is then output to the classification output layer. The sequence attention weight is enhanced at the sequence position corresponding to the end of the feeding process to focus on the discrimination of interlayer transition events.

[0010] Furthermore, the pre-training process of the material pushing state discrimination model includes: Constructing a training dataset: The torque sequence and encoder sequence in the historical pusher data are converted into the standard stroke torque sequence according to the adaptive stroke reconstruction step; wherein, the interlayer transition state label is accurately calibrated based on the falling edge of the torque curve and the pusher position information; Design loss function: Use weighted cross-entropy loss function for training, where the weight assigned to the inter-layer transition state category is higher than the weight assigned to the empty push state and payload state categories; Model optimization: Using the standard stroke torque sequence as input and its corresponding state label as supervision signal, the parameters of the temporal convolutional neural network are optimized using the backpropagation algorithm to minimize the loss function.

[0011] Furthermore, before making a final control decision based on the confidence level satisfying the first preset condition or the second preset condition, a first-level physical verification step is also included: When it is determined that the material box may be empty or that interlayer transition may occur, a physical verification index is calculated. For determining that the material box is empty, the physical verification index is the average value of the load energy integral of the current and the previous N consecutive material pushing actions; For the determination of interlayer transition, the physical verification index is the load energy integral value of the current pushing action; The load energy integral is obtained by integrating the displacement over the effective travel interval of the real-time torque sequence; The physical verification indicators are compared with the corresponding energy threshold ranges; the energy threshold ranges are obtained based on product specifications and historical normal material feeding data. The final stop loop or inter-layer switching operation will only be executed when the confidence level condition and the corresponding physical verification index condition are met simultaneously; if the two contradict each other, an alarm will be triggered or a preset review and security process will be executed.

[0012] Furthermore, the first preset condition includes: In M consecutive pushing actions, during the latter half of the pushing stroke of each action, the percentage of time during which the confidence level of the empty pushing state continuously exceeds the first dynamic threshold exceeds the set proportion.

[0013] Furthermore, the second preset condition is a logical combination of the following conditions: The confidence level of the interlayer transition state exceeds the second threshold; The pusher head's real-time stroke position is within the preset end-stage push range; The real-time torque sequence was detected to be in a downward trend.

[0014] Furthermore, during the single-layer pushing action of the pushing device, the output torque sampling value of the servo motor driving the pusher head is collected in real time to form a real-time torque sequence, specifically including: With a fixed sampling period, the output torque value of the servo motor and the motor encoder value reflecting the position of the push head are synchronously triggered and collected; The acquired raw output torque value is digitally filtered in real time to suppress high-frequency switching noise and power frequency interference; Based on the trend of the filtered torque value and the encoder value, the start and end points of a single pushing action are detected, and all synchronously collected data points are extracted to form the real-time torque sequence and its corresponding synchronous position sequence.

[0015] The present invention also includes a real-time material pushing status determination system based on torque sensing, for a material pushing device that pushes products out of the material box layer by layer, the system comprising: The torque data acquisition module is used to acquire the output torque sampling value of the servo motor driving the pusher head in real time during the single-layer pusher action of the pusher, and form a real-time torque sequence. The material pushing state discrimination module is used to input the torque sequence into a pre-trained material pushing state discrimination model. The material pushing state discrimination model is a temporal convolutional neural network, which extracts dynamic features at different time scales from the real-time torque sequence synchronously through stacked dilated causal convolutional layers, and outputs the confidence scores for at least three material pushing states, including the empty pushing state, the effective load state, and the interlayer transition state. The decision control module is used to make decisions based on the confidence level. When the confidence level of the empty push state meets the first preset condition, it determines that the material box is empty and controls the push mechanism to stop the cycle. When the confidence level of the interlayer transition state meets the second preset condition, it triggers the push mechanism to perform an interlayer switching operation.

[0016] The beneficial effects of this invention are as follows: by acquiring the output torque sequence of the servo motor in real time, extracting dynamic features and outputting the confidence level of the pushing state using a pre-trained temporal convolutional neural network model, and making decision control based on the confidence level, the invention can accurately identify the empty pushing state and the interlayer transition state, ensuring the pushing accuracy and reliability. It has the advantages of real-time and accurate judgment of the pushing state, avoiding misjudgment of empty pushing and interlayer transition delay, improving equipment reliability and production efficiency, and reducing fatigue damage to mechanical parts.

[0017] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the material feeding device; Figure 2 The flowchart is a method for real-time determination of material feeding status based on torque sensing. Figure 3 A flowchart illustrating the adaptive trip reconstruction steps; Figure 4 A flowchart illustrating the attention fusion layer configured to execute steps; Figure 5 This is a schematic diagram of a real-time material feeding status discrimination system based on torque sensing. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] Example 1: like Figures 1 to 4 As shown, this application provides a real-time material pushing state determination method based on torque sensing, used in a material pushing device to push products out of a material box layer by layer. The method includes: S10: During the process of the pushing equipment performing a single-layer pushing action, the output torque sampling value of the servo motor driving the push head is collected in real time to form a real-time torque sequence; S20: Input the torque sequence into the pre-trained push state discrimination model; the push state discrimination model is a temporal convolutional neural network, which extracts dynamic features at different time scales from the real-time torque sequence synchronously through stacked dilated causal convolutional layers, and outputs the confidence of at least three push states, including empty push state, effective load state and inter-layer transition state. S30: Make decisions based on confidence level. When the confidence level of the empty push state meets the first preset condition, determine that the material box is empty and control the push mechanism to stop the cycle. When the confidence level of the interlayer transition state meets the second preset condition, trigger the push mechanism to perform the interlayer switching operation.

[0023] This method is applied to a feeding device that pushes products out of a material box layer by layer. The feeding device includes a feeding mechanism, a servo motor that drives the pusher head, a controller, and a feeding status determination module that communicates with the controller; for example... Figure 1As shown, in this embodiment, the pushing mechanism on the left is used to perform a single-layer pushing action under the control command, that is, the pusher head completes a complete push and return motion along the predetermined pushing direction to push out the product located in the current layer in the middle material box as a whole. Specifically, during the single-layer pushing action of the pushing device, the controller collects the sampled value of the output torque of the servo motor driving the pusher head in real time through the servo driver. The torque sampled value can be the electromagnetic torque estimate calculated internally by the servo driver, or it can be obtained by conversion from current feedback. The acquisition method is used to obtain torque information that can reflect the force change during the pushing process. The real-time torque sequence is then input into the pre-trained pushing state discrimination model to realize the real-time identification of the current pushing state. The pushing state discrimination model is a temporal convolutional neural network, whose network structure is composed of multiple dilated causal convolutional layers stacked in time order. Each dilated causal convolutional layer adopts a causal convolution method that only depends on the current and historical input, thereby ensuring the temporal causality of the output of the pushing state discrimination model. To ensure the temporal consistency of the model during the material pushing process and prevent future information leakage, the aforementioned material pushing state discrimination model structure outputs a confidence distribution corresponding to each of the above-mentioned material pushing states based on the input real-time torque sequence. Then, after obtaining the confidence scores for each state output by the material pushing state discrimination model, the controller executes corresponding control operations based on preset decision logic. Based on this confidence score, the controller can make corresponding decisions: when the confidence score of the empty pushing state exceeds a preset threshold, the controller will stop the material pushing cycle; when the confidence score of the inter-layer transition state meets the conditions, the controller triggers the material pushing mechanism to perform an inter-layer switching operation, thereby allowing the material pushing process to continue according to a predetermined rhythm. Furthermore, the training process of the material pushing state discrimination model involves collecting labeled torque sequence samples, enabling the model to possess strong generalization ability. Through this method, the material pushing equipment can accurately identify the material pushing state and automatically adjust its working state without relying on additional sensors, improving the intelligence level, operational reliability, and product delivery consistency of the material pushing equipment.

[0024] Firstly, by acquiring real-time torque data from the motor output and analyzing it using a temporal convolutional neural network, the limitations of traditional methods relying on fixed thresholds are effectively avoided. Unlike methods based on position or vision sensors, torque sensing technology can directly reflect dynamic force changes during the feeding process, adapting in real-time to various changing conditions in production, such as product tolerances, mechanical wear, speed adjustments, and friction changes, thus improving the robustness and adaptability of feeding state identification. Secondly, the application of the temporal convolutional neural network enables accurate identification of the complex and transient torque change process during interlayer transitions, avoiding misjudgments or confusion caused by relying on a single threshold point in traditional technologies. By extracting dynamic features at different time scales, the feeding state can be accurately identified. Subtle changes ensure precise control of interlayer switching timing, improve production cycle accuracy, and reduce the risk of equipment interference. Furthermore, utilizing torque data instead of relying on traditional photoelectric or position sensors reduces the system's dependence on the external environment, avoiding interference from factors such as lighting and surface morphology of materials. This approach is relatively low-cost and requires no complex hardware deployment or maintenance, thus improving the equipment's economy and maintainability. In short, real-time torque sensing and intelligent judgment overcome the limitations of traditional material feeding status judgment methods, providing high-precision and high-reliability material feeding status judgment in dynamic and complex industrial environments. This effectively improves the intelligence and automation level of the material feeding system, optimizes production efficiency, and extends equipment lifespan.

[0025] As a preferred embodiment of the above, in step S20, before inputting the real-time torque sequence into the pre-trained material pushing state discrimination model, such as... Figure 3 As shown, it also includes an adaptive route reconstruction step: S21: Synchronously acquire the servo motor encoder signal at the moment corresponding to the output torque sampling value; S22: Determine the starting moment of this pushing action based on the real-time torque sequence; S23: From the real-time torque sequence, detect the starting point where the torque value first decreases from the stable load platform, and record it as the torque decoupling point; S24: Take the encoder position corresponding to the torque release point as the start of the stroke, and take the encoder position corresponding to the torque release point as the effective physical end of the stroke. S25: Based on the actual displacement between the start of the journey and the end of the effective physical journey, normalization is performed to obtain a normalized journey sequence; S26: Based on the normalized stroke sequence, the real-time torque sequence is resampled and interpolated to generate a standard stroke torque sequence of fixed length; The standard stroke torque sequence serves as the input to the material pushing state discrimination model.

[0026] Specifically, during the operation of the feeding device, the servo motor encoder is used to monitor the position information of the pusher head in real time. At the moment corresponding to the collected motor torque value, the servo motor encoder signal must be collected synchronously. This is to ensure that the torque data is synchronized with the physical stroke data, providing accurate reference information for subsequent stroke reconstruction. Typically, the servo motor encoder signal can be directly connected to the control system through the encoder's output interface. The collected encoder signal is stored and processed together with the real-time motor torque sampling value. Determining the starting moment of this feeding action is also crucial. Typically, the starting point of the feeding action is the instant the motor begins to apply thrust. In practical applications, by analyzing the real-time torque sequence, changes in torque values ​​can be observed. This marks the start of the pushing action. At this point, the control system identifies the starting moment of pushing based on the changing trend of the torque sequence. Generally, the starting moment of pushing coincides with the instant when the torque begins to rise sharply from a low value. Then, during the pushing process, the torque release point refers to the moment when the contact between the pusher head and the product begins to weaken and the product is about to be ejected from the material box, at which point the torque value first decreases. This change marks the end of the current layer of pushing and prepares for the next layer of pushing. To accurately identify this point, the fluctuation characteristics in the real-time torque sequence are analyzed, especially the starting point of the first decrease in torque value from the stable load platform, i.e., the torque value under effective load conditions. Once the torque release point is determined, the encoder position corresponding to that torque release point is used as the starting point of the push. The effective physical stroke endpoint is determined by using the encoder position corresponding to the start of the pushing action as the starting point of the effective physical stroke. This method clearly defines the product's movement range throughout the pushing process, providing an accurate data foundation for subsequent stroke normalization and reconstruction. Then, the actual displacement data of the pushing action—the displacement from the stroke start point to the effective physical stroke endpoint—is used as the normalization benchmark. Normalization eliminates physical differences caused by equipment or environmental factors, allowing all pushing data to be compared and processed using a unified standard. The normalized stroke sequence ensures data consistency and comparability between different pushing tasks. Finally, the real-time acquired torque sequence is re-processed... Sampling and interpolation processes generate a standard stroke torque sequence of fixed length. This standard stroke torque sequence serves as the input data for the subsequent material feeding state discrimination model. Resampling and interpolation ensure the consistency of the length and time scale of the torque sequence, avoiding inconsistencies in input data caused by stroke differences in different material feeding tasks. Through the above adaptive stroke reconstruction steps, the relationship between torque and physical stroke can be accurately identified and synchronized during actual material feeding, thereby providing high-quality, standardized input data and further improving the accuracy and reliability of material feeding state discrimination. Compared with traditional methods, this scheme can effectively cope with the impact of equipment status, environmental changes, and differences in mechanical performance, ensuring efficient and precise control of the material feeding process.

[0027] In this embodiment, the temporal convolutional neural network includes an input layer, a multi-path feature extraction module, and a classification output layer connected in sequence; The multi-path feature extraction module includes a transient feature extraction subnetwork, a trend feature extraction subnetwork, and a frequency domain feature extraction subnetwork, all configured in parallel. The transient feature extraction subnetwork includes causal convolutional layers with a first set of dilation coefficients configured to have a receptive field on a millisecond time scale for extracting torque abrupt features characterizing the pusher contacting the product and the product leaving the hopper from a standard stroke torque sequence. The trend feature extraction subnetwork includes a causal convolutional layer with a second set of expansion coefficients, which are greater than the first set of expansion coefficients. This layer is configured to have a receptive field that covers the main stroke timescale of a single push, and is used to extract load plateau and trend features that characterize the stable push process of the entire layer of product from the standard stroke torque sequence. The frequency domain feature extraction subnetwork includes a time-frequency transform layer and subsequent convolutional layers, which are used to extract frequency domain energy distribution features that characterize the sliding friction vibration and abnormal impact of the product from the transform domain of the standard stroke torque sequence.

[0028] Specifically, by introducing a multi-path feature extraction module, the traditional method for determining the material pushing state is improved, enabling more accurate and robust monitoring of the material pushing process. This module includes transient, trend, and frequency domain feature extraction sub-networks, allowing the model to process torque changes at different time scales and with different signal characteristics during the material pushing process. The transient feature extraction sub-network focuses on capturing torque abrupt changes, particularly at the moments when the pusher contacts the product and the product leaves the hopper, identifying rapidly changing torque signals. The trend feature extraction sub-network focuses on identifying stable material pushing processes over long time scales, extracting load platform and trend features to help determine whether the material pushing process is stable. The frequency domain feature extraction sub-network extracts frequency domain energy features related to frictional vibration and abnormal impacts through time-frequency transformation and convolution operations, further improving the accuracy of material pushing state determination. Through these multi-path feature processing methods, the accuracy and reliability of the material pushing process are improved. Especially in complex and dynamic industrial environments, it can accurately identify different stages of the material pushing state, avoiding misjudgment problems in traditional methods and enhancing the intelligent and automated control capabilities of the equipment.

[0029] The temporal convolutional neural network also includes an attention fusion layer located between the multi-path feature extraction module and the classification output layer; such as Figure 4 As shown, the attention fusion layer is configured to perform the following operations: A21: The feature maps output by the transient feature extraction subnetwork, the trend feature extraction subnetwork, and the frequency domain feature extraction subnetwork are concatenated along the channel dimension to obtain a concatenated feature map sequence; A22: The normalized travel sequence is mapped to a travel position encoding vector that matches the length of the spliced ​​feature map sequence through a sine and cosine position encoding function; A23: Add the concatenated feature map to the run position encoding vector, and generate channel attention weights and sequence attention weights respectively through a two-layer feedforward neural network; A24: The concatenated feature map is weighted and fused using channel attention weights and sequence attention weights to obtain a fused feature map, which is then output to the classification output layer. The sequence attention weights are enhanced at the sequence positions corresponding to the end of the feeding process to focus on the discrimination of interlayer transition events.

[0030] Specifically, firstly, in the multi-path feature extraction module, transient features, trend features, and frequency domain features are extracted independently through their respective sub-networks. Each sub-network generates a feature map representing different features extracted from the standard stroke torque sequence. The first step of the attention fusion layer is to concatenate these feature maps along the channel dimension, thereby integrating all extracted feature information into a unified feature map. Next, the stroke sequence, representing the actual displacement data during the material pushing process, is normalized, typically within the range of 0 to 1. To incorporate stroke information into the feature map, we use a sine and cosine position encoding function to map the normalized stroke sequence into a stroke position encoding vector. The purpose of this step is... Introducing location information into the time series allows the network to better understand and utilize the travel location information at each time step. Specifically, the location encoding function encodes the normalized travel value at each time step using sine and cosine functions, generating a fixed-length encoding vector, which is then fused with the feature map. Next, the concatenated feature map is added to the travel location encoding vector to obtain a fused feature map. This fused feature map is then input into a two-layer feedforward neural network. This network generates channel attention weights and sequence attention weights. Channel attention weights: the network learns the importance of different channels in the feature map and assigns different weights to each channel, which helps... The network focuses on more important feature channels, employing sequence attention weights. These weights focus on different positions in the input sequence, especially towards the end of the feeding process, i.e., the inter-layer transition phase, strengthening the discriminative ability at this stage. Finally, channel attention weights and sequence attention weights are applied to the concatenated feature map for weighted fusion. Specifically, first, the features of each channel are weighted using channel attention weights, and then the position of the feature map in the entire sequence is weighted using sequence attention weights, ultimately resulting in a fused feature map. This fused feature map is then passed as input to the classification output layer to generate the final feeding state discrimination result. This weighted fusion process ensures that important information during the feeding process is preserved. In particular, inter-layer transition events can receive sufficient attention, allowing for more accurate determination of the feeding state, especially at critical moments in the feeding process, ensuring the accuracy and timeliness of state switching. By introducing an attention fusion layer, the accuracy of feeding state determination can be effectively improved, especially at critical moments of inter-layer transition. The attention mechanism enables the network to automatically focus on the most important features and sequence positions, thereby reducing interference from irrelevant information. Through the weighted fusion of channel attention and sequence attention, more accurate and robust feeding state determination can be provided in dynamic and complex industrial environments, improving the level of intelligent control of the feeding process and optimizing production efficiency.

[0031] In this embodiment, the pre-training process of the material pushing state discrimination model includes: Constructing a training dataset: The torque sequence and encoder sequence in the historical pusher data are converted into a standard stroke torque sequence according to the adaptive stroke reconstruction step; among them, the interlayer transition state label is accurately calibrated based on the falling edge of the torque curve and the pusher position information; Loss function design: The weighted cross-entropy loss function is used for training, where the weights assigned to the inter-layer transition state category are higher than the weights assigned to the empty push state and payload state categories. Model optimization: Using the standard stroke torque sequence as input and its corresponding state label as supervision signal, the backpropagation algorithm is used to optimize the parameters of the temporal convolutional neural network to minimize the loss function.

[0032] Specifically, firstly, the training dataset is generated by collecting torque sequences from historical pusher data and synchronous encoder sequences, and then generating a standard stroke torque sequence in a unified format according to the adaptive stroke reconstruction steps described above. In data annotation, the transient and critical category of interlayer transition states is accurately labeled based on the significant falling edge of the torque curve and combined with pusher position change information to ensure label accuracy. To address the potential imbalance of pusher state category samples (such as fewer interlayer transition samples), the training uses a weighted cross-entropy loss function. By assigning higher class weights to interlayer transition states, the model focuses more on recognizing this critical state during the learning process. In the model optimization stage, the standard stroke torque sequence is used as input, and the corresponding accurate state labels are used as supervision signals. The parameters of the temporal convolutional neural network are continuously adjusted using the backpropagation algorithm to minimize the weighted loss function, ultimately resulting in a robust model with high discrimination accuracy.

[0033] As a preferred embodiment of the above, step S30 further includes a first-level physical verification step before making a final control decision based on the confidence level meeting the first preset condition or the second preset condition: S31: When it is determined that the material box may be empty or interlayer transition may occur, calculate a physical verification index; For determining whether the material box is empty, the physical verification index is the average value of the load energy integral of the current and the previous N consecutive material pushing actions. For interlayer transition determination, the physical verification index is the load energy integral value of the current pushing action; The load energy integral is obtained by integrating the displacement over the real-time torque sequence within the effective travel range; S32: Compare the physical verification indicators with the corresponding energy threshold range; the energy threshold range is obtained based on product specifications and historical normal material feeding data. S33: The final stop loop or inter-layer switching operation will only be executed when the confidence condition and the corresponding physical verification index condition are met simultaneously; if the two contradict each other, an alarm will be triggered or a preset review and security process will be executed.

[0034] Specifically, when the confidence level output by the material feeding status discrimination model meets the preset conditions that the material box may be empty or that interlayer transition may occur, the loop stop or interlayer switching operation is not immediately executed. Instead, the first-level physical verification process is initiated. In this embodiment, when the model determines that the material box may be empty, the calculated physical verification index is the average load energy integral of the current and the previous N consecutive material feeding actions. Here, N can be set according to the equipment cycle time and stability requirements. By introducing the average value of multiple material feedings, abnormal judgments caused by occasional disturbances or noise during a single material feeding process can be effectively suppressed. When the model determines that interlayer transition... When this occurs, the calculated physical verification index is the integral value of the load energy of the current pushing action. This is because interlayer transitions typically exhibit a rapid decay of load energy at the end of a single pushing stroke, exhibiting a distinct single-stroke characteristic. The load energy integral is obtained by integrating the real-time torque sequence along the pusher displacement direction within the effective stroke interval. Specifically, within the determined effective physical stroke interval, the torque value at each sampling moment is multiplied by the corresponding small displacement and accumulated to obtain the integral value of the load energy representing the actual work done during this pushing process. Then, the calculated physical verification index is compared with the pre-defined... The set energy threshold range is compared. This energy threshold range is not a fixed empirical value, but is obtained based on product specification parameters, such as product weight, size, friction characteristics, and historical normal feeding data through statistical analysis. After completing the above physical verification, the confidence condition output by the model is jointly judged with the physical verification index condition. Only when both simultaneously meet the corresponding judgment condition will the final control operation be executed, such as stopping the feeding cycle or triggering inter-layer switching. If the model confidence judgment result contradicts the physical verification result, for example, the model determines that the material box is empty but the load energy integral is still higher than the empty feeding threshold, the control operation will be terminated. Within the specified range, instead of performing direct control actions, an alarm will be triggered, or a preset review and safety process will be initiated, such as reducing the material pushing speed, performing a trial material pushing, or requesting manual confirmation, thereby avoiding production risks caused by accidental shutdowns or switching. As can be seen from the above implementation method, by introducing a first-level physical verification step based on load energy integration, the discrimination results of the deep learning model are combined with physically interpretable quantities, which effectively improves the reliability and safety of material pushing status discrimination in complex industrial environments, especially in critical material box emptying and interlayer transition discrimination scenarios, reducing the risk of misjudgment and erroneous actions.

[0035] In step S30, the first preset condition includes: In M consecutive pushing actions, the percentage of time during the latter half of the pushing stroke of each action when the confidence level of the empty pushing state continuously exceeds the first dynamic threshold exceeds the set proportion.

[0036] Specifically, the first preset condition improves the reliability of the judgment by introducing a comprehensive criterion of consistency of multiple pushes, constraints in the second half of the stroke, and the proportion of confidence time. In each push action, the push state discrimination model outputs the confidence sequence of the empty push state along the normalized push stroke, dividing a single push stroke into the first half and the second half, and taking the second half of the push stroke as the key discrimination interval. In this interval, the proportion of time that the confidence of the empty push state continuously exceeds the first dynamic threshold is statistically analyzed. The first dynamic threshold can be adaptively adjusted based on historical push data or current working conditions. On this basis, a judgment mechanism for M consecutive push actions is introduced. Only when the proportion of time that the confidence of the empty push state exceeds the first dynamic threshold in the second half of the stroke of each push action in the M consecutive pushes is higher than the set proportion, is the first preset condition that the material box is empty considered to be met. Through the above method, misjudgments caused by single push abnormalities, instantaneous fluctuations, or model uncertainties are effectively avoided, making the material box emptying judgment more stable and reliable, and providing a credible basis for subsequent physical verification and final control decisions.

[0037] In this embodiment, in step S30, the second preset condition is a logical combination of the following conditions: The confidence level of the interlayer transition state exceeds the second threshold; The pusher head's real-time stroke position is within the preset end-stage push range; The real-time torque sequence was detected to be in a downward trend.

[0038] Specifically, the second preset condition was designed in combination, combining the model's judgment result with the pusher stroke information and torque change trend for joint constraint. Specifically, during a single pusher stroke, it not only judges whether the confidence level of the interlayer transition state exceeds the second threshold, but also simultaneously detects whether the real-time stroke position of the pusher head is within the preset end-stage pusher interval, and analyzes whether the real-time torque sequence shows a downward trend. The end-stage pusher interval is defined based on the normalized stroke sequence and is used to limit the interlayer transition judgment to occur in a physically reasonable pusher stage. The torque downward trend is used to reflect the actual mechanical process of the product gradually leaving the material box and the pusher head load decreasing. Only when the above three conditions are met simultaneously is the interlayer transition state determined to be established and the interlayer switching operation is triggered, thereby effectively avoiding misjudgments caused by instantaneous confidence fluctuations, abnormal friction, or non-end-stage strokes, and improving the accuracy and stability of the interlayer switching timing determination.

[0039] As a preferred embodiment of the above, in step S10, during the single-layer pushing action of the pushing device, the output torque sampling value of the servo motor driving the pusher head is collected in real time to form a real-time torque sequence, specifically including: S11: Synchronously trigger and collect the output torque value of the servo motor and the motor encoder value reflecting the position of the push head with a fixed sampling period; S12: Perform real-time digital filtering on the acquired raw output torque value to suppress high-frequency switching noise and power frequency interference; S13: Based on the trend of the filtered torque value change and the encoder value, detect the start and end points of a single pushing action, and extract all synchronously collected data points in between to form a real-time torque sequence and its corresponding synchronous position sequence.

[0040] Specifically, firstly, the data acquisition interface of the servo driver is synchronously triggered with a preset fixed sampling period to obtain the real-time output torque value of the servo motor and the motor encoder value reflecting the pusher position information. By adopting a synchronous triggering method, it is ensured that each torque sampling point corresponds one-to-one with the unique pusher position data, thus providing a reliable foundation for subsequent displacement-based stroke analysis and torque integration. Next, real-time digital filtering is performed on the raw output torque value acquired in real time to suppress high-frequency switching noise and power frequency interference commonly found in servo drive systems. In a specific embodiment, low-pass digital filtering, moving average filtering, or band-stop filtering can be used to process the torque signal online, so that the filtered torque signal retains the true force change trend while effectively removing high-frequency jitter unrelated to the pushing physical process. This step avoids noise signals being misidentified as effective torque changes in subsequent start and end point detection and state judgment. Finally, based on the trend of the filtered torque value change and the same... The encoder values ​​collected step by step are used to automatically detect the start and end points of a single pushing action, and all synchronously collected data points are extracted during this period. In one embodiment, when the torque starts to show a continuous upward trend from a low and stable state, and the encoder position begins to change significantly, this moment is determined as the start point of a single pushing action; when the torque decreases significantly and tends to stabilize after passing through the load platform, and the encoder position change slows down or stops, this moment is determined as the end point of the pushing action; subsequently, only all synchronously sampled data between the start and end points are retained to form a real-time torque sequence and its synchronous position sequence corresponding to a single pushing action. Through the above implementation method, data extraction for a single pushing action can be adaptively completed without relying on external trigger signals or fixed time windows. This avoids the introduction of invalid data and ensures the integrity of key force and displacement information throughout the pushing process, providing a high-quality data input foundation for subsequent stroke reconstruction, feature extraction, and state discrimination.

[0041] Example 2: This invention also includes a real-time material pushing status determination system based on torque sensing, used for a material pushing device that pushes products out of the material box layer by layer, such as... Figure 5 As shown, the system includes: The torque data acquisition module is used to collect the output torque sampling value of the servo motor driving the pusher head in real time during the single-layer pusher action of the pusher equipment, and form a real-time torque sequence. The material pushing state discrimination module is used to input the torque sequence into the pre-trained material pushing state discrimination model. The material pushing state discrimination model is a temporal convolutional neural network. Through stacked dilated causal convolutional layers, it synchronously extracts dynamic features at different time scales from the real-time torque sequence and outputs the confidence of at least three material pushing states, including empty pushing state, effective load state and inter-layer transition state. The decision control module is used to make decisions based on confidence level. When the confidence level of the empty push state meets the first preset condition, it determines that the material box is empty and controls the push mechanism to stop the cycle. When the confidence level of the interlayer transition state meets the second preset condition, it triggers the push mechanism to perform the interlayer switching operation.

[0042] The adjustment system described above in this invention can effectively realize the real-time determination method of the feeding state based on torque sensing, and the technical effects it can achieve are as described in the above embodiments, and will not be repeated here.

[0043] Similarly, the above-mentioned optimization schemes for the system can also achieve the optimization effects corresponding to the methods in Embodiment 1, which will not be repeated here.

[0044] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and accompanying drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for real-time determination of material feeding status based on torque sensing, characterized in that, A feeding device for pushing products out of a material box layer by layer, the method comprising: During the process of the pushing device performing a single-layer pushing action, the output torque sampling value of the servo motor driving the push head is collected in real time to form a real-time torque sequence; The torque sequence is input into a pre-trained push state discrimination model; the push state discrimination model is a temporal convolutional neural network, which synchronously extracts dynamic features at different time scales from the real-time torque sequence through stacked dilated causal convolutional layers, and outputs the confidence of at least three push states, including empty push state, effective load state and inter-layer transition state. Decisions are made based on the confidence level. When the confidence level of the empty push state meets the first preset condition, it is determined that the material box is empty and the push mechanism is controlled to stop the cycle. When the confidence level of the interlayer transition state meets the second preset condition, the push mechanism is triggered to perform an interlayer switching operation.

2. The real-time material feeding state determination method based on torque sensing according to claim 1, characterized in that, Before inputting the real-time torque sequence into the pre-trained material pushing state discrimination model, an adaptive stroke reconstruction step is also included: The servo motor encoder signal is synchronously acquired at the moment corresponding to the output torque sampling value; Based on the real-time torque sequence, the starting time of this pushing action is determined; From the real-time torque sequence, the starting point at which the torque value first decreases from the stable load platform is detected is recorded as the torque decoupling point; The encoder position corresponding to the torque release point is taken as the start point of the stroke, and the encoder position corresponding to the torque release point is taken as the effective physical end point of the stroke. Based on the actual displacement between the starting point of the journey and the end point of the effective physical journey, normalization processing is performed to obtain a normalized journey sequence; Based on the normalized stroke sequence, the real-time torque sequence is resampled and interpolated to generate a standard stroke torque sequence of fixed length; The standard stroke torque sequence serves as the input to the material pushing state discrimination model.

3. The real-time material feeding state determination method based on torque sensing according to claim 2, characterized in that, The temporal convolutional neural network includes an input layer, a multi-path feature extraction module, and a classification output layer connected in sequence. The multi-path feature extraction module includes a transient feature extraction sub-network, a trend feature extraction sub-network, and a frequency domain feature extraction sub-network that are configured in parallel. The transient feature extraction subnetwork includes a causal convolutional layer with a first set of dilation coefficients configured to have a receptive field on a millisecond time scale, for extracting torque abrupt features characterizing the pusher contacting the product and the product leaving the hopper from the standard stroke torque sequence. The trend feature extraction subnetwork includes a causal convolutional layer with a second set of expansion coefficients, which are greater than the first set of expansion coefficients. It is configured to have a receptive field that covers the main stroke time scale of a single push, and is used to extract load plateau and trend features that characterize the stable push process of the entire layer of product from the standard stroke torque sequence. The frequency domain feature extraction subnetwork includes a time-frequency transform layer and subsequent convolutional layers, used to extract frequency domain energy distribution features characterizing the sliding friction vibration and abnormal impact of the product from the transform domain of the standard stroke torque sequence.

4. The real-time material feeding state determination method based on torque sensing according to claim 3, characterized in that, The temporal convolutional neural network further includes an attention fusion layer located between the multi-path feature extraction module and the classification output layer, the attention fusion layer being configured to perform the following operations: The feature maps output by the transient feature extraction subnetwork, the trend feature extraction subnetwork, and the frequency domain feature extraction subnetwork are concatenated along the channel dimension to obtain a concatenated feature map sequence. The normalized travel sequence is mapped to a travel position encoding vector that matches the length of the spliced ​​feature map sequence using a sine-cosine position encoding function; The concatenated feature map is added to the travel position encoding vector, and then a two-layer feedforward neural network is used to generate channel attention weights and sequence attention weights, respectively. The channel attention weight and sequence attention weight are applied to the spliced ​​feature map for weighted fusion to obtain a fused feature map, which is then output to the classification output layer. The sequence attention weight is enhanced at the sequence position corresponding to the end of the feeding process to focus on the discrimination of interlayer transition events.

5. The real-time material feeding state determination method based on torque sensing according to claim 2, characterized in that, The pre-training process of the material pushing state discrimination model includes: Constructing a training dataset: The torque sequence and encoder sequence in the historical pusher data are converted into the standard stroke torque sequence according to the adaptive stroke reconstruction step; wherein, the interlayer transition state label is accurately calibrated based on the falling edge of the torque curve and the pusher position information; Design loss function: Use weighted cross-entropy loss function for training, where the weight assigned to the inter-layer transition state category is higher than the weight assigned to the empty push state and payload state categories; Model optimization: Using the standard stroke torque sequence as input and its corresponding state label as supervision signal, the parameters of the temporal convolutional neural network are optimized using the backpropagation algorithm to minimize the loss function.

6. The real-time material feeding state determination method based on torque sensing according to claim 1, characterized in that, Before making a final control decision based on the confidence level meeting the first or second preset condition, a first-level physical verification step is also included: When it is determined that the material box may be empty or that interlayer transition may occur, a physical verification index is calculated. For determining that the material box is empty, the physical verification index is the average value of the load energy integral of the current and the previous N consecutive material pushing actions; For the determination of interlayer transition, the physical verification index is the load energy integral value of the current pushing action; The load energy integral is obtained by integrating the displacement over the effective travel interval of the real-time torque sequence; The physical verification indicators are compared with the corresponding energy threshold ranges; the energy threshold ranges are obtained based on product specifications and historical normal material feeding data. The final stop loop or inter-layer switching operation will only be executed when the confidence level condition and the corresponding physical verification index condition are met simultaneously; if the two contradict each other, an alarm will be triggered or a preset review and security process will be executed.

7. The real-time material feeding state determination method based on torque sensing according to claim 1, characterized in that, The first preset conditions include: In M consecutive pushing actions, during the latter half of the pushing stroke of each action, the percentage of time during which the confidence level of the empty pushing state continuously exceeds the first dynamic threshold exceeds the set proportion.

8. The real-time material feeding state determination method based on torque sensing according to claim 1, characterized in that, The second preset condition is a logical combination of the following conditions: The confidence level of the interlayer transition state exceeds the second threshold; The pusher head's real-time stroke position is within the preset end-stage push range; The real-time torque sequence was detected to be in a downward trend.

9. The real-time material feeding state determination method based on torque sensing according to claim 1, characterized in that, During the single-layer pushing action of the pushing device, the output torque sampling value of the servo motor driving the pusher head is collected in real time to form a real-time torque sequence, specifically including: With a fixed sampling period, the output torque value of the servo motor and the motor encoder value reflecting the position of the push head are synchronously triggered and collected; The acquired raw output torque value is digitally filtered in real time to suppress high-frequency switching noise and power frequency interference; Based on the trend of the filtered torque value and the encoder value, the start and end points of a single pushing action are detected, and all synchronously collected data points are extracted to form the real-time torque sequence and its corresponding synchronous position sequence.

10. A real-time material feeding state discrimination system based on torque sensing, characterized in that, A feeding device for pushing products out of a material box layer by layer, the system comprising: The torque data acquisition module is used to acquire the output torque sampling value of the servo motor driving the pusher head in real time during the single-layer pusher action of the pusher, and form a real-time torque sequence. The material pushing state discrimination module is used to input the torque sequence into a pre-trained material pushing state discrimination model. The material pushing state discrimination model is a temporal convolutional neural network, which extracts dynamic features at different time scales from the real-time torque sequence synchronously through stacked dilated causal convolutional layers, and outputs the confidence scores for at least three material pushing states, including the empty pushing state, the effective load state, and the interlayer transition state. The decision control module is used to make decisions based on the confidence level. When the confidence level of the empty push state meets the first preset condition, it determines that the material box is empty and controls the push mechanism to stop the cycle. When the confidence level of the interlayer transition state meets the second preset condition, it triggers the push mechanism to perform an interlayer switching operation.

Images