A filling machine material error prediction system based on multi-sensor fusion

Abstract

The application discloses a kind of based on multi-sensor fusion's filling machine material error prediction system, comprising: multi-sensor data acquisition and fusion module, for collecting multi-sensor data set and working condition data set, constructs multi-sensor fusion input data sequence;Deterministic filling quantity prediction module, for generating reference filling quantity prediction sequence;Material filling error state construction module, for constructing material filling error state sequence;Improved SDE-Net modeling module, for constructing improved SDE-Net model, configuration model parameter;Quantization diffusion modeling module, for constructing error random disturbance structure parameter;Working condition synergistic modulation module, for generating synergistic modulation parameter and executing synchronous modulation update;Material filling error prediction module, for executing continuous time evolution prediction processing and generating filling machine material error prediction result.The application improves the accuracy and stability of filling machine material error prediction.

Description

Technical Field

[0001] This invention relates to the field of packaging machinery and filling equipment technology, and in particular to a material error prediction system for filling machines based on multi-sensor fusion. Background Technology

[0002] As a key piece of equipment in the packaging production line, the filling machine's filling accuracy directly affects product quality and production stability. In actual production, due to changes in material properties, fluctuations in equipment operating conditions, and external environmental factors, discrepancies often occur between the actual filling volume and the target filling volume, resulting in filling errors. How to effectively model and predict filling errors in advance has always been a key technical issue in the filling equipment field.

[0003] In existing technologies, the handling of filling errors mostly focuses on post-processing detection or simple statistical analysis. These methods typically rely on empirical judgments based on single sensor data or a limited number of operating parameters, making it difficult to comprehensively reflect the error evolution under the coupling effects of multiple physical quantities during the filling process. While some solutions incorporate multi-sensor data, they often remain at the level of feature stitching or static modeling, lacking the characterization of the continuous changes in error over time and failing to adapt to dynamic fluctuations under complex operating conditions.

[0004] Furthermore, existing filling error prediction methods often fail to distinguish between deterministic changes and random disturbance sources in error evolution, and lack a structured modeling mechanism for the impact of disturbances introduced by different sensors, resulting in insufficient stability of prediction results. At the same time, the influence of operating conditions on error evolution paths and disturbance intensity is usually reflected in simple condition input methods, making it difficult to achieve a coordinated characterization of multiple operating states.

[0005] Therefore, how to provide a material error prediction system for filling machines based on multi-sensor fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] This invention proposes a material error prediction system for filling machines based on multi-sensor fusion. By acquiring multi-sensor data sets and operating condition data sets, a multi-sensor fusion input data sequence is constructed to generate a baseline filling volume prediction sequence and a material filling error state sequence. Based on this, an improved SDE-Net model is built to model the deterministic evolution parameters and random perturbation structure parameters of the error. Furthermore, a fractional diffusion structure and an operating condition collaborative modulation mechanism are introduced to achieve continuous-time prediction of the material filling error state. This invention models the error as an evolvable state, which can characterize the comprehensive impact of multi-source perturbations and operating condition changes on the evolution of filling errors, thus improving the accuracy and stability of material error prediction for filling machines.

[0007] A material error prediction system for a filling machine based on multi-sensor fusion according to an embodiment of the present invention includes:

[0008] The multi-sensor data acquisition and fusion module is used to acquire multi-sensor data sets and operating condition data sets, perform time synchronization processing and data alignment processing, and construct a multi-sensor fused input data sequence.

[0009] The deterministic fill volume prediction module is used to receive multi-sensor fusion input data sequences, perform deterministic modeling processing through the deterministic fill volume prediction substructure, and generate a baseline fill volume prediction sequence.

[0010] The material filling error state construction module is used to generate an actual filling quantity sequence based on measurement data, perform difference calculation processing with the benchmark filling quantity prediction sequence, and construct a material filling error state sequence.

[0011] An improved SDE-Net modeling module is used to construct an improved SDE-Net model based on the material filling error state sequence. Drift subnetwork and diffusion subnetwork are set in the model to generate a set of error deterministic evolution parameters and channel disturbance contribution parameters.

[0012] The fractional diffusion modeling module is used to construct the fractional diffusion structure in the diffusion subnetwork and construct the error random perturbation structure parameters.

[0013] The working condition cooperative modulation module is used to construct a working condition representation vector based on the working condition data set, generate cooperative modulation parameters, perform synchronous modulation update, and obtain the modulated error deterministic evolution parameters and the modulated error random perturbation structure parameters.

[0014] The material filling error prediction module is used to perform continuous-time evolution prediction processing in the improved SDE-Net model to generate material error prediction results for the filling machine.

[0015] Optionally, modules can be integrated using the following methods:

[0016] Acquire the multi-sensor data set generated during the operation of the filling machine, construct the multi-sensor fusion input data sequence, and simultaneously acquire the operating condition data set;

[0017] The multi-sensor fusion input data sequence is input into the deterministic filling volume prediction substructure, and deterministic modeling processing is performed on the filling process to generate a baseline filling volume prediction sequence.

[0018] The actual filling volume sequence is obtained based on the multi-sensor data set, and the difference calculation is performed between it and the baseline filling volume prediction sequence to generate a material filling error state sequence.

[0019] An improved SDE-Net model is constructed based on the material filling error state sequence, including a drift subnetwork and a diffusion subnetwork. It receives the material filling error state sequence and generates a set of error deterministic evolution parameters and channel perturbation contribution parameters.

[0020] In the diffusion subnetwork, a fractional diffusion structure is constructed based on the set of channel perturbation contribution parameters, and a set of multi-sensor data is accepted to construct the error random perturbation structure parameters;

[0021] Based on the set of operating condition data, an operating condition representation vector is constructed. The drift subnetwork and the diffusion subnetwork are input, and parameter co-modulation processing is performed to generate co-modulation parameters and perform parameter modulation update.

[0022] In the improved SDE-Net model, the material filling error state prediction calculation is performed based on the result of parameter modulation update, and the material error prediction result of the filling machine is generated.

[0023] Optionally, the generation of the multi-sensor fusion input data sequence and the operating condition data set includes:

[0024] The raw data collected by each sensor during the operation of the filling machine are processed to generate a multi-sensor data set. The pressure data, flow data, vibration data, temperature data and metering data in the multi-sensor data set are recorded according to the corresponding collection timestamps. The metering data represents the actual filling volume within the corresponding filling cycle.

[0025] Perform time synchronization and data alignment processing on the multi-sensor dataset and organize them jointly to construct a multi-sensor fused input data sequence.

[0026] The system collects and processes data on valve opening, target filling volume, filling cycle time, and material batch during the operation of the filling machine to generate a set of operating condition data.

[0027] Optionally, the generation of the baseline fill volume prediction sequence includes:

[0028] The multi-sensor fused input data sequence is input into the deterministic filling volume prediction substructure. An input organization unit is set in the deterministic filling volume prediction substructure, joint organization processing is performed, and a deterministic input feature sequence is constructed.

[0029] A deterministic feature mapping unit is set up to receive a deterministic input feature sequence, perform deterministic computation processing, and generate an intermediate representation for filling quantity prediction.

[0030] Set up a filling volume output unit to receive the intermediate representation of the filling volume prediction, perform numerical mapping processing, and organize it into a baseline filling volume prediction sequence.

[0031] Optionally, the generation of the material filling error state sequence includes:

[0032] Based on the metering data in the multi-sensor dataset, the metering data is organized and processed according to the filling cycle marker to obtain the actual filling volume sequence;

[0033] Based on the baseline filling volume prediction sequence and the actual filling volume sequence, the filling volume error value corresponding to each filling cycle is calculated and arranged to generate a filling volume error time series;

[0034] Based on the filling volume error time series, the filling volume error values ​​of adjacent filling cycles are processed to form a temporal continuity organization and construct an error state representation.

[0035] Based on the error state representation, an error state vector is constructed, which includes the filling volume error value corresponding to the current filling cycle, the error change between adjacent filling cycles, and the cumulative change of the filling volume error.

[0036] The error state vector is combined with the filling cycle mark and the material batch mark to construct a material filling error state sequence.

[0037] Optionally, the generation of the set of error deterministic evolution parameters and channel perturbation contribution parameters includes:

[0038] The improved SDE-Net model structure is constructed based on the material filling error state sequence, and a drift subnetwork and a diffusion subnetwork are constructed in the state space of the model.

[0039] The material filling error state sequence is input into the drift sub-network, and state mapping processing is performed based on the current state of the error state sequence to generate error deterministic evolution parameters.

[0040] The material filling error state sequence and the multi-sensor fusion input data sequence are jointly input into the diffusion sub-network. Based on the joint features of the error state and the multi-sensor fusion input data, disturbance modeling processing is performed to generate a set of channel disturbance contribution parameters.

[0041] Optionally, the construction and use of the fractionation diffusion structure includes:

[0042] In the diffusion subnetwork, a fractional diffusion structure is constructed based on the set of channel disturbance contribution parameters, and pressure diffusion channel, flow diffusion channel, vibration diffusion channel and temperature diffusion channel are set.

[0043] Pressure data from the multi-sensor fusion input data sequence is input into the pressure diffusion channel, flow data into the flow diffusion channel, vibration data into the vibration diffusion channel, and temperature data into the temperature diffusion channel. Disturbance modeling is performed in each diffusion channel based on the change characteristics of the corresponding physical quantity, generating pressure disturbance contribution parameters, flow disturbance contribution parameters, vibration disturbance contribution parameters, and temperature disturbance contribution parameters, respectively.

[0044] A unified scaling process is performed on the pressure disturbance contribution parameters, flow disturbance contribution parameters, vibration disturbance contribution parameters, and temperature disturbance contribution parameters;

[0045] Based on the unified scale mapping process, according to the correspondence between the error state dimension and each perturbation contribution parameter, weighted integration and dimension alignment are performed to map each perturbation contribution parameter to the same random perturbation structure representation and construct the error random perturbation structure parameter.

[0046] By using the random perturbation structure parameters of the error as the perturbation input term for the random evolution of the error state in the improved SDE-Net model, the random perturbation generated by the fractional diffusion structure can directly participate in the random evolution modeling of the material filling error state.

[0047] Optionally, the parameter modulation update includes:

[0048] Based on the working condition data set, the working condition elements are jointly organized and numerically encoded according to the filling cycle to construct a working condition representation vector.

[0049] The working condition representation vector is jointly mapped to the material filling error state sequence, and the collaborative modulation parameters are generated based on the joint mapping result.

[0050] The cooperative modulation parameters are input into the output modulation position of the drift subnetwork and the perturbation aggregation modulation position of the diffusion subnetwork, respectively, and act on the generation process of the error deterministic evolution parameters and the aggregation process of the error random perturbation structure parameters.

[0051] Within the same filling cycle, the deterministic error evolution parameters and the random error perturbation structure parameters are synchronously modulated and updated to obtain the modulated deterministic error evolution parameters and the modulated random error perturbation structure parameters.

[0052] Optionally, the generation of the material error prediction result for the filling machine includes:

[0053] The improved SDE-Net model is jointly input with the multi-sensor fusion input data sequence, the working condition representation vector and the material filling error state sequence. Based on the constructed drift subnetwork and diffusion subnetwork, the continuous time advancement start state of error state prediction is determined.

[0054] Under the constraints of the modulated error deterministic evolution parameters, the material filling error state sequence is subjected to continuous-time evolution progression processing to obtain the deterministic evolution result of the error state;

[0055] Under the constraints of the modulated error random perturbation structure parameters, the continuous-time evolution process of the error state is superimposed with random perturbation propagation processing to generate the error state evolution result;

[0056] The error state evolution result is used as the material filling error state input for the next time step. Multi-time step continuous prediction processing is performed according to the preset time advancement rules to generate a material filling error prediction sequence, including the error state prediction value corresponding to each prediction time step. The material error prediction result of the filling machine is determined based on the error state prediction value at the corresponding target prediction time.

[0057] The beneficial effects of this invention are:

[0058] First, this invention constructs a multi-sensor fusion input data sequence and forms a material filling error state sequence based on it, thereby elevating the filling error from a single static deviation to a state variable that can continuously evolve over time. This allows the material error of the filling machine to be continuously characterized during operation, effectively overcoming the problem in existing technologies that rely solely on post-event statistics or discrete error analysis and are unable to reflect dynamic changes.

[0059] Secondly, this invention introduces an improved SDE-Net model in the error modeling process, which decomposes the error evolution process into two key elements: the deterministic evolution parameters of the error and the structural parameters of the random disturbance of the error. It also performs structured modeling of the disturbance effects corresponding to pressure data, flow data, vibration data and temperature data through fractional diffusion structure, so as to more accurately reflect the actual effect of multi-source sensing information on filling error fluctuations and improve the stability and reliability of error prediction results under complex working conditions.

[0060] Furthermore, this invention constructs a working condition representation vector based on a set of working condition data, and synchronously modulates the deterministic evolution parameters of the error and the random disturbance structure parameters of the error through collaborative modulation parameters. This ensures that changes in working conditions have a consistent impact on the error evolution path and the intensity of random disturbances, thereby enhancing the model's adaptability to different filling cycles and material batch conditions. It enables continuous time prediction of material errors in filling machines, improving overall prediction accuracy and application value. Attached Figure Description

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0062] Fig. 1This is an overall flowchart of a filling machine material error prediction system based on multi-sensor fusion proposed in this invention;

[0063] Fig. 2 This is a schematic diagram of the structure of the improved SDE-Net model in this invention;

[0064] Fig. 3 This is a schematic diagram of the structure of the fractional diffusion modeling and working condition co-modulation mechanism in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figs. 1-3 A material error prediction system for filling machines based on multi-sensor fusion, comprising:

[0067] The multi-sensor data acquisition and fusion module is used to collect multi-sensor data sets and operating condition data sets during the operation of the filling machine. The multi-sensor data set includes pressure data, flow data, vibration data, temperature data, and metering data. The metering data is used to characterize the actual filling volume. The operating condition data set includes valve opening data, target filling volume set value, filling cycle time identifier, and material batch identifier. Time synchronization processing and data alignment processing are performed on the multi-sensor data set, and a multi-sensor fusion input data sequence is constructed based on the results of time synchronization processing and data alignment processing to provide input data with a unified time granularity for subsequent processing modules.

[0068] The deterministic filling volume prediction module is used to receive multi-sensor fusion input data sequences, set up a deterministic filling volume prediction substructure inside the module, perform deterministic modeling processing on the multi-sensor fusion input data sequences, and generate a benchmark filling volume prediction sequence that corresponds one-to-one with the filling cycle.

[0069] The material filling error state construction module is used to generate the actual filling volume sequence based on the metering data in the multi-sensor data set, and to perform difference calculation processing based on the actual filling volume sequence and the benchmark filling volume prediction sequence to construct a material filling error state sequence that reflects the continuous change of filling volume error with the filling cycle, which serves as the state input for subsequent stochastic evolution modeling.

[0070] An improved SDE-Net modeling module is used to construct an improved SDE-Net model based on the material filling error state sequence. In the model, a drift subnetwork and a diffusion subnetwork are set. The drift subnetwork is used to generate the deterministic evolution parameters of the error, and the diffusion subnetwork is used to generate the set of channel disturbance contribution parameters. The deterministic evolution modeling and random disturbance modeling of the error state are completed in the model state space.

[0071] The fractional diffusion modeling module is used to construct a fractional diffusion structure based on the set of channel perturbation contribution parameters in the diffusion subnetwork. In the fractional diffusion structure, pressure diffusion channels, flow diffusion channels, vibration diffusion channels and temperature diffusion channels are set. Perturbation modeling processing is performed on pressure data, flow data, vibration data and temperature data respectively to generate pressure perturbation contribution parameters, flow perturbation contribution parameters, vibration perturbation contribution parameters and temperature perturbation contribution parameters. Error random perturbation structure parameters are constructed based on unified scale mapping, weighted integration and dimension alignment processing.

[0072] The working condition co-modulation module is used to construct a working condition representation vector based on the working condition data set, perform joint mapping processing on the working condition representation vector and the material filling error state sequence to generate co-modulation parameters, and apply the co-modulation parameters to the output modulation position of the drift sub-network and the perturbation aggregation modulation position of the diffusion sub-network respectively. Within the same filling cycle, the error deterministic evolution parameters and the error random perturbation structure parameters are synchronously modulated and updated to obtain the modulated error deterministic evolution parameters and the modulated error random perturbation structure parameters.

[0073] The material filling error prediction module is used to input the multi-sensor fusion input data sequence, the working condition representation vector, and the material filling error state sequence into the improved SDE-Net model. Under the constraints of the modulated error deterministic evolution parameters and the modulated error random perturbation structure parameters, it performs continuous time evolution advancement and random perturbation advancement processing on the material filling error state, generates a material filling error prediction sequence arranged in time sequence, and determines the material error prediction result of the filling machine based on the error state prediction value at the corresponding target prediction time in the material filling error prediction sequence.

[0074] In this embodiment, the modules are interconnected using the following method:

[0075] The system acquires a multi-sensor data set and a working condition data set generated during the operation of the filling machine. The multi-sensor data set includes pressure data, flow data, vibration data, temperature data, and metering data. The metering data represents the actual filling volume. The working condition data set includes valve opening data, target filling volume set value, filling cycle time identifier, and material batch identifier. The system performs time synchronization processing and data alignment processing on the multi-sensor data set and constructs a multi-sensor fusion input data sequence based on the results of the time synchronization processing and data alignment processing.

[0076] The multi-sensor fusion input data sequence is input into the deterministic filling volume prediction substructure. Based on the deterministic filling volume prediction substructure, deterministic modeling processing is performed on the filling process to generate a baseline filling volume prediction sequence.

[0077] The actual filling volume sequence is obtained based on the measurement data in the multi-sensor dataset. The difference between the actual filling volume sequence and the benchmark filling volume prediction sequence is calculated to generate the material filling error state sequence. The material filling error state sequence is then used as the state input of the improved SDE-Net model.

[0078] An improved SDE-Net model is constructed based on the material filling error state sequence. The improved SDE-Net model includes a drift subnetwork and a diffusion subnetwork. The drift subnetwork receives the material filling error state sequence and generates error deterministic evolution parameters. The diffusion subnetwork receives the material filling error state sequence and the multi-sensor fusion input data sequence and generates a set of channel perturbation contribution parameters.

[0079] In the diffusion subnetwork, a fractional diffusion structure is constructed based on the set of channel disturbance contribution parameters. The fractional diffusion structure includes a pressure diffusion channel, a flow diffusion channel, a vibration diffusion channel, and a temperature diffusion channel. The pressure diffusion channel receives pressure data and outputs pressure disturbance contribution parameters, the flow diffusion channel receives flow data and outputs flow disturbance contribution parameters, the vibration diffusion channel receives vibration data and outputs vibration disturbance contribution parameters, and the temperature diffusion channel receives temperature data and outputs temperature disturbance contribution parameters. Based on the pressure disturbance contribution parameters, flow disturbance contribution parameters, vibration disturbance contribution parameters, and temperature disturbance contribution parameters, error random disturbance structure parameters are constructed.

[0080] A working condition representation vector is constructed based on the working condition data set. The working condition representation vector is input into the drift subnetwork and the diffusion subnetwork. Parameter co-modulation processing is performed to generate co-modulation parameters. Based on the co-modulation parameters, parameter modulation is performed on the error deterministic evolution parameters to obtain the modulated error deterministic evolution parameters. Based on the co-modulation parameters, parameter modulation is performed on the error random perturbation structure parameters to obtain the modulated error random perturbation structure parameters.

[0081] The improved SDE-Net model inputs multi-sensor fusion input data sequence, operating condition representation vector, and material filling error state sequence. Based on the modulated error deterministic evolution parameters and the modulated error random perturbation structure parameters, the material filling error state prediction calculation is performed to generate the material error prediction result for the filling machine.

[0082] In this embodiment, the generation of the multi-sensor fusion input data sequence and the operating condition data set includes:

[0083] The raw data collected by each sensor during the operation of the filling machine are processed to generate a multi-sensor data set. The pressure data, flow data, vibration data, temperature data and metering data in the multi-sensor data set are recorded according to the corresponding collection timestamps. The metering data is used to characterize the actual filling volume within the corresponding filling cycle.

[0084] Time synchronization processing is performed on the multi-sensor data set. Based on a unified filling cycle identifier, the pressure data, flow data, vibration data, temperature data, and metering data are aligned on the time axis to generate a time-synchronized multi-sensor data sequence.

[0085] Data alignment processing is performed on the time-synchronized multi-sensor data sequence. According to the preset time granularity, various sensor data are resampled and interpolated to form one-to-one corresponding data items for pressure data, flow data, vibration data, temperature data and measurement data at the same time granularity.

[0086] Based on the completion of time synchronization and data alignment processing, pressure data, flow data, vibration data, temperature data and measurement data at the same time granularity are jointly organized to construct a multi-sensor fusion input data sequence;

[0087] The system collects and processes data on valve opening, target filling volume, filling cycle time, and material batch during the operation of the filling machine, generating a set of operating condition data. This set of operating condition data is then stored in conjunction with the multi-sensor fusion input data sequence for time correlation, so that it can be called upon in subsequent processing steps.

[0088] In this embodiment, the generation of the baseline filling volume prediction sequence includes:

[0089] The multi-sensor fusion input data sequence is input into the deterministic filling volume prediction substructure. An input organization unit is set in the deterministic filling volume prediction substructure. The input organization unit performs joint organization processing on the pressure data, flow data, vibration data, temperature data and metering data in the multi-sensor fusion input data sequence according to the filling cycle, and constructs a deterministic input feature sequence that corresponds one-to-one with the filling cycle.

[0090] A deterministic feature mapping unit is set in the deterministic filling volume prediction substructure. The deterministic feature mapping unit receives a deterministic input feature sequence and performs deterministic computation processing on the deterministic input feature sequence based on a mapping structure with fixed parameters to generate an intermediate representation of filling volume prediction used to characterize the relationship between multi-sensor input features and filling volume. The deterministic computation processing does not introduce random variables, does not perform probability sampling, and does not generate uncertainty parameters.

[0091] A filling quantity output unit is set in the deterministic filling quantity prediction substructure. The filling quantity output unit receives the intermediate representation of the filling quantity prediction, performs numerical mapping processing on the intermediate representation of the filling quantity prediction, generates the benchmark filling quantity prediction value corresponding to each filling cycle, and organizes it into a benchmark filling quantity prediction sequence according to the filling cycle order.

[0092] In this embodiment, the generation of the material filling error state sequence includes:

[0093] Based on the metering data in the multi-sensor dataset, the metering data is organized and processed according to the filling cycle marker to obtain the actual filling volume sequence that corresponds one-to-one with each filling cycle.

[0094] Based on the baseline filling volume prediction sequence and the actual filling volume sequence, the difference between the two is calculated according to the filling cycle to obtain the filling volume error value corresponding to each filling cycle. The filling volume error values ​​are then arranged in order of the filling cycle to generate a filling volume error time series.

[0095] Based on the filling volume error time series, the filling volume error values ​​of adjacent filling cycles are processed by time continuity organization to construct an error state representation that reflects the continuous change of filling volume error with the filling cycle.

[0096] Based on the error state representation, an error state vector is constructed, which includes the filling quantity error value corresponding to the current filling cycle, the error change between adjacent filling cycles, and the cumulative change of the filling quantity error. This vector is used to characterize the evolution characteristics of the filling quantity error in the time dimension.

[0097] The error state vector is jointly organized with the filling cycle mark and the material batch mark to construct a material filling error state sequence associated with the filling cycle mark and the material batch.

[0098] The material filling error state sequence is used as the state input of the improved SDE-Net model for subsequent material filling error prediction processing based on stochastic evolution.

[0099] In this embodiment, the generation of the error deterministic evolution parameter set and the channel perturbation contribution parameter set includes:

[0100] The model structure of the improved SDE-Net model is constructed based on the material filling error state sequence. The improved SDE-Net model sets up a state space to describe the error state evolution, and constructs a drift subnetwork and a diffusion subnetwork in the state space of the model. The drift subnetwork and the diffusion subnetwork serve as the deterministic branch and the random perturbation branch of the error evolution in the state space, respectively.

[0101] The material filling error state sequence is input into the drift sub-network. In the drift sub-network, state mapping processing is performed based on the current state of the error state sequence to generate error deterministic evolution parameters that characterize the evolution trend of material filling error over time.

[0102] The material filling error state sequence and the multi-sensor fusion input data sequence are jointly input into the diffusion sub-network. In the diffusion sub-network, perturbation modeling is performed based on the joint features of the error state and the multi-sensor fusion input data to generate a set of channel perturbation contribution parameters to characterize the influence of perturbations in different sensor channels.

[0103] In the improved SDE-Net model, the set of error deterministic evolution parameters and channel disturbance contribution parameters are used as deterministic inputs and random disturbance inputs for subsequent error state stochastic evolution modeling, respectively, for use in subsequent material filling error prediction processing.

[0104] In this embodiment, the construction and use of the fractionation diffusion structure includes:

[0105] In the diffusion subnetwork, a fractional diffusion structure is constructed based on the set of channel disturbance contribution parameters. The fractional diffusion structure sets up pressure diffusion channels, flow diffusion channels, vibration diffusion channels and temperature diffusion channels in the random disturbance modeling path of the diffusion subnetwork. Each diffusion channel corresponds to a disturbance component from a different physical source.

[0106] Pressure data from the multi-sensor fusion input data sequence is input into the pressure diffusion channel, flow data into the flow diffusion channel, vibration data into the vibration diffusion channel, and temperature data into the temperature diffusion channel. Disturbance modeling is performed in each diffusion channel based on the change characteristics of the corresponding physical quantity, generating pressure disturbance contribution parameters, flow disturbance contribution parameters, vibration disturbance contribution parameters, and temperature disturbance contribution parameters, respectively.

[0107] In the fractional diffusion structure, a unified scale mapping process is performed on the pressure disturbance contribution parameters, flow disturbance contribution parameters, vibration disturbance contribution parameters and temperature disturbance contribution parameters, so that the disturbance contribution parameters from different physical sources have a consistent numerical scale and comparability in the same random disturbance modeling space.

[0108] Based on the unified scale mapping process, according to the correspondence between the error state dimension and each disturbance contribution parameter, the pressure disturbance contribution parameter, flow disturbance contribution parameter, vibration disturbance contribution parameter and temperature disturbance contribution parameter are weighted and integrated and dimension aligned. Each disturbance contribution parameter is mapped to the same random disturbance structure representation, and error random disturbance structure parameters are constructed to characterize the comprehensive influence of multi-source disturbances.

[0109] By using the random perturbation structure parameters of the error as the perturbation input term for the random evolution of the error state in the improved SDE-Net model, the random perturbation generated by the fractional diffusion structure can directly participate in the random evolution modeling of the material filling error state.

[0110] In this embodiment, the parameter modulation update includes:

[0111] Based on the valve opening data, target filling volume setting, filling cycle mark and material batch mark in the working condition data set, the working condition elements are jointly organized and numerically encoded according to the filling cycle to construct a working condition representation vector to characterize the operating status of the filling machine.

[0112] The working condition representation vector and the material filling error state sequence are jointly mapped, and the co-modulation parameters used to characterize the impact of working condition changes on the error state evolution are generated based on the joint mapping results.

[0113] The cooperative modulation parameters are input into the output modulation position of the drift subnetwork and the perturbation aggregation modulation position of the diffusion subnetwork, respectively. In the drift subnetwork, the cooperative modulation parameters are applied to the generation process of the error deterministic evolution parameters, and in the diffusion subnetwork, the cooperative modulation parameters are applied to the aggregation process of the error random perturbation structure parameters.

[0114] Within the same filling cycle, the deterministic evolution parameters of the error and the random disturbance structure parameters of the error are synchronously modulated and updated so that the working condition representation vector produces a consistent modulation effect on the deterministic evolution path of the error state and the random disturbance intensity, thus obtaining the modulated deterministic evolution parameters of the error and the modulated random disturbance structure parameters of the error.

[0115] In this embodiment, the generation of the material error prediction result for the filling machine includes:

[0116] The multi-sensor fusion input data sequence, the working condition representation vector, and the material filling error state sequence are input into the improved SDE-Net model. Based on the drift subnetwork and diffusion subnetwork already constructed in the improved SDE-Net model, the continuous time advancement start state of error state prediction is determined.

[0117] Under the constraints of the modulated error deterministic evolution parameters, the material filling error state sequence is subjected to continuous-time evolution progression processing to obtain the deterministic evolution result of the error state;

[0118] Under the constraints of the modulated error random perturbation structure parameters, the continuous-time evolution process of the error state is superimposed with random perturbation propagation processing to generate an error state evolution result containing the influence of random perturbation.

[0119] The error state evolution result, which includes the influence of random disturbances, is used as the material filling error state input for the next time step. Multi-time step continuous prediction processing is performed according to the preset time advancement rules to generate a material filling error prediction sequence arranged in time order. The material filling error prediction sequence includes the error state prediction value corresponding to each prediction time step. The material error prediction result of the filling machine is determined based on the error state prediction value corresponding to the target prediction time in the material filling error prediction sequence.

[0120] Example 1:

[0121] To verify the feasibility of this invention in practice, it was applied to a continuous liquid filling production scenario to predict and analyze the material filling errors generated by the filling machine during actual operation. In this scenario, the filling machine operates under continuous start-stop and variable load conditions for extended periods. The material is affected by various factors during its flow, including pressure fluctuations, flow rate fluctuations, mechanical vibration, temperature changes, and switching of operating parameters, leading to a certain deviation between the actual filling volume and the target filling volume. Traditional filling error control methods often rely on single metering feedback or simple empirical models, making it difficult to predict the temporal evolution trend of errors in advance. Especially when the filling cycle time changes or material batches are switched, error fluctuations exhibit significant randomness and cumulativeity, easily leading to a decline in filling consistency.

[0122] In this scenario, pressure, flow, vibration, temperature, and metering data generated during the filling machine's operation are collected uniformly, and a set of operational data consisting of valve opening, target filling volume setpoint, filling cycle time identifier, and material batch identifier is acquired simultaneously. Through time synchronization and data alignment processing, a multi-sensor fusion input data sequence is constructed, ensuring continuous input of various data types along the same time axis. Based on this multi-sensor fusion input data sequence, the deterministic filling volume prediction module performs deterministic modeling of the filling process, generating a baseline filling volume prediction sequence reflecting the ideal filling state. Subsequently, the difference between the actual filling volume sequence formed by metering data and the baseline filling volume prediction sequence is calculated to construct a material filling error state sequence, transforming the filling error from a single numerical value into a state variable that evolves continuously over time.

[0123] Building upon this foundation, an improved SDE-Net model is constructed based on the material filling error state sequence. Within the model, a drift subnetwork and a diffusion subnetwork are set up to model the deterministic evolution trend and random disturbance characteristics of the error, respectively. A fractional diffusion modeling module is introduced into the diffusion subnetwork, allowing pressure, flow, vibration, and temperature data to correspond to independent diffusion channels, generating disturbance contribution parameters with physically meaningful mapping relationships, and constructing random disturbance structure parameters. A working condition representation vector is constructed by combining the working condition data set, and a collaborative modulation module generates collaborative modulation parameters. Synchronous modulation updates are performed on the deterministic error evolution parameters and the random disturbance structure parameters, enabling the model to reflect changes in error evolution characteristics under different filling cycle times and material batch conditions. Finally, continuous-time evolution prediction processing is performed in the improved SDE-Net model to generate material error prediction results for the filling machine, used to assess the filling error change trend within future time windows.

[0124] During continuous operation testing, data from multiple filling cycles were selected as evaluation samples to compare the error prediction performance of the method described in this invention with that of traditional methods based on single metering feedback. The test data covered both stable operating conditions and frequent switching conditions, reflecting the error variation characteristics in a real production environment. Statistical results focused on multi-time-step error prediction, emphasizing the magnitude of the predicted error, the stability of error fluctuations, and the ability to respond to sudden error changes.

[0125] Table 1 Comparison of the performance of different methods in predicting filling errors

[0126]

[0127] As can be seen from the data in Table 1, under the same test conditions, the method of this invention reduces the mean absolute prediction error by approximately 0.7 grams compared to the traditional method. The prediction results are closer to the actual filling error level, and the maximum prediction error is effectively controlled, indicating that the present invention has a better ability to characterize abnormal changes when the error fluctuation is large. The decrease in the standard deviation of the error indicates that the prediction results are more stable in the time dimension, reducing the phenomenon of large jumps in the predicted values. In terms of the multi-time step prediction stability index, the method of this invention shows higher consistency and can maintain the continuity of the error trend during continuous prediction. In the filling cycle switching stage, the prediction error of the method of this invention is significantly smaller than that of the traditional method, indicating that through the working condition collaborative modulation mechanism, the model can better adapt to the impact of changes in operating conditions.

[0128] This invention, through the synergistic effect of multi-sensor fusion, material filling error state modeling, and improved SDE-Net structure, enables filling error prediction to no longer be limited to single feedback correction, but to reflect the evolution law of error in continuous time. The fractional diffusion modeling and working condition collaborative modulation mechanism enable random disturbances and operating conditions to be reasonably characterized, ensuring improved prediction accuracy.

[0129] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A material error prediction system for filling machines based on multi-sensor fusion, characterized in that, Includes the following steps: The multi-sensor data acquisition and fusion module is used to acquire multi-sensor data sets and operating condition data sets, perform time synchronization processing and data alignment processing, and construct a multi-sensor fused input data sequence. The deterministic fill volume prediction module is used to receive multi-sensor fusion input data sequences, perform deterministic modeling processing through the deterministic fill volume prediction substructure, and generate a baseline fill volume prediction sequence. The material filling error state construction module is used to generate an actual filling quantity sequence based on measurement data, perform difference calculation processing with the benchmark filling quantity prediction sequence, and construct a material filling error state sequence. An improved SDE-Net modeling module is used to construct an improved SDE-Net model based on the material filling error state sequence. Drift subnetwork and diffusion subnetwork are set in the model to generate a set of error deterministic evolution parameters and channel disturbance contribution parameters. The fractional diffusion modeling module is used to construct the fractional diffusion structure in the diffusion subnetwork and construct the error random perturbation structure parameters. The working condition cooperative modulation module is used to construct a working condition representation vector based on the working condition data set, generate cooperative modulation parameters, perform synchronous modulation update, and obtain the modulated error deterministic evolution parameters and the modulated error random perturbation structure parameters. The material filling error prediction module is used to perform continuous-time evolution prediction processing in the improved SDE-Net model to generate material error prediction results for the filling machine.

2. The filling machine material error prediction system based on multi-sensor fusion according to claim 1, characterized in that, The modules are connected in the following way: Acquire the multi-sensor data set generated during the operation of the filling machine, construct the multi-sensor fusion input data sequence, and simultaneously acquire the operating condition data set; The multi-sensor fusion input data sequence is input into the deterministic filling volume prediction substructure, and deterministic modeling processing is performed on the filling process to generate a baseline filling volume prediction sequence. The actual filling volume sequence is obtained based on the multi-sensor data set, and the difference calculation is performed between it and the baseline filling volume prediction sequence to generate a material filling error state sequence. An improved SDE-Net model is constructed based on the material filling error state sequence, including a drift subnetwork and a diffusion subnetwork. It receives the material filling error state sequence and generates a set of error deterministic evolution parameters and channel perturbation contribution parameters. In the diffusion subnetwork, a fractional diffusion structure is constructed based on the set of channel perturbation contribution parameters, and a set of multi-sensor data is accepted to construct the error random perturbation structure parameters; Based on the set of operating condition data, an operating condition representation vector is constructed. The drift subnetwork and the diffusion subnetwork are input, and parameter co-modulation processing is performed to generate co-modulation parameters and perform parameter modulation update. In the improved SDE-Net model, the material filling error state prediction calculation is performed based on the result of parameter modulation update, and the material error prediction result of the filling machine is generated.

3. The filling machine material error prediction system based on multi-sensor fusion according to claim 2, characterized in that, The generation of the multi-sensor fusion input data sequence and operating condition data set includes: The raw data collected by each sensor during the operation of the filling machine are processed to generate a multi-sensor data set. The pressure data, flow data, vibration data, temperature data and metering data in the multi-sensor data set are recorded according to the corresponding collection timestamps. The metering data represents the actual filling volume within the corresponding filling cycle. Perform time synchronization and data alignment processing on the multi-sensor dataset and organize them jointly to construct a multi-sensor fused input data sequence. The system collects and processes data on valve opening, target filling volume, filling cycle time, and material batch during the operation of the filling machine to generate a set of operating condition data.

4. The filling machine material error prediction system based on multi-sensor fusion according to claim 2, characterized in that, The generation of the baseline fill volume prediction sequence includes: The multi-sensor fused input data sequence is input into the deterministic filling volume prediction substructure. An input organization unit is set in the deterministic filling volume prediction substructure, joint organization processing is performed, and a deterministic input feature sequence is constructed. A deterministic feature mapping unit is set up to receive a deterministic input feature sequence, perform deterministic computation processing, and generate an intermediate representation for filling quantity prediction. Set up a filling volume output unit to receive the intermediate representation of the filling volume prediction, perform numerical mapping processing, and organize it into a baseline filling volume prediction sequence.

5. The filling machine material error prediction system based on multi-sensor fusion according to claim 2, characterized in that, The generation of the material filling error state sequence includes: Based on the metering data in the multi-sensor dataset, the metering data is organized and processed according to the filling cycle marker to obtain the actual filling volume sequence; Based on the baseline filling volume prediction sequence and the actual filling volume sequence, the filling volume error value corresponding to each filling cycle is calculated and arranged to generate a filling volume error time series; Based on the filling volume error time series, the filling volume error values ​​of adjacent filling cycles are processed to form a temporal continuity organization and construct an error state representation. Based on the error state representation, an error state vector is constructed, which includes the filling volume error value corresponding to the current filling cycle, the error change between adjacent filling cycles, and the cumulative change of the filling volume error. The error state vector is combined with the filling cycle mark and the material batch mark to construct a material filling error state sequence.

6. The filling machine material error prediction system based on multi-sensor fusion according to claim 2, characterized in that, The generation of the set of error deterministic evolution parameters and channel perturbation contribution parameters includes: The improved SDE-Net model structure is constructed based on the material filling error state sequence, and a drift subnetwork and a diffusion subnetwork are constructed in the state space of the model. The material filling error state sequence is input into the drift sub-network, and state mapping processing is performed based on the current state of the error state sequence to generate error deterministic evolution parameters. The material filling error state sequence and the multi-sensor fusion input data sequence are jointly input into the diffusion sub-network. Based on the joint features of the error state and the multi-sensor fusion input data, disturbance modeling processing is performed to generate a set of channel disturbance contribution parameters.

7. The filling machine material error prediction system based on multi-sensor fusion according to claim 2, characterized in that, The construction and use of the fractionation diffusion structure include: In the diffusion subnetwork, a fractional diffusion structure is constructed based on the set of channel disturbance contribution parameters, and pressure diffusion channel, flow diffusion channel, vibration diffusion channel and temperature diffusion channel are set. Pressure data from the multi-sensor fusion input data sequence is input into the pressure diffusion channel, flow data into the flow diffusion channel, vibration data into the vibration diffusion channel, and temperature data into the temperature diffusion channel. Disturbance modeling is performed in each diffusion channel based on the change characteristics of the corresponding physical quantity, generating pressure disturbance contribution parameters, flow disturbance contribution parameters, vibration disturbance contribution parameters, and temperature disturbance contribution parameters, respectively. A unified scaling process is performed on the pressure disturbance contribution parameters, flow disturbance contribution parameters, vibration disturbance contribution parameters, and temperature disturbance contribution parameters; Based on the unified scale mapping process, according to the correspondence between the error state dimension and each perturbation contribution parameter, weighted integration and dimension alignment are performed to map each perturbation contribution parameter to the same random perturbation structure representation and construct the error random perturbation structure parameter. By using the random perturbation structure parameters of the error as the perturbation input term for the random evolution of the error state in the improved SDE-Net model, the random perturbation generated by the fractional diffusion structure can directly participate in the random evolution modeling of the material filling error state.

8. The filling machine material error prediction system based on multi-sensor fusion according to claim 2, characterized in that, The parameter modulation update includes: Based on the working condition data set, the working condition elements are jointly organized and numerically encoded according to the filling cycle to construct a working condition representation vector. The working condition representation vector is jointly mapped to the material filling error state sequence, and the collaborative modulation parameters are generated based on the joint mapping result. The cooperative modulation parameters are input into the output modulation position of the drift subnetwork and the perturbation aggregation modulation position of the diffusion subnetwork, respectively, and act on the generation process of the error deterministic evolution parameters and the aggregation process of the error random perturbation structure parameters. Within the same filling cycle, the deterministic error evolution parameters and the random error perturbation structure parameters are synchronously modulated and updated to obtain the modulated deterministic error evolution parameters and the modulated random error perturbation structure parameters.

9. A material error prediction system for a filling machine based on multi-sensor fusion according to claim 2, characterized in that, The generation of the material error prediction results for the filling machine includes: The improved SDE-Net model is jointly input with the multi-sensor fusion input data sequence, the working condition representation vector and the material filling error state sequence. Based on the constructed drift subnetwork and diffusion subnetwork, the continuous time advancement start state of error state prediction is determined. Under the constraints of the modulated error deterministic evolution parameters, the material filling error state sequence is subjected to continuous-time evolution progression processing to obtain the deterministic evolution result of the error state; Under the constraints of the modulated error random perturbation structure parameters, the continuous-time evolution process of the error state is superimposed with random perturbation propagation processing to generate the error state evolution result; The error state evolution result is used as the material filling error state input for the next time step. Multi-time step continuous prediction processing is performed according to the preset time advancement rules to generate a material filling error prediction sequence, including the error state prediction value corresponding to each prediction time step. The material error prediction result of the filling machine is determined based on the error state prediction value at the corresponding target prediction time.

Images