Remote data acquisition control system for spray product filling production line

Abstract

The application discloses a remote data acquisition control system for a spray product filling production line, and relates to the technical field of industrial automation control. The remote data acquisition control system comprises a data acquisition module, a feature fusion module, a target matching module and a parameter optimization module. The data acquisition module is used for collecting multi-source heterogeneous data of the spray product filling production line based on a chaos mapping initialized particle swarm population. The feature fusion module is used for performing feature fusion on the multi-source heterogeneous data through an adaptive fuzzy neural network to generate a state feature vector of the production line. The target matching module is used for performing similarity matching on the state feature vector and a preset filling process parameter template to determine a regulation and control target of a current production stage. The parameter optimization module is used for iteratively optimizing regulation and control parameters through a chaos particle swarm algorithm to generate a remote control instruction. The instruction sending module is used for sending the remote control instruction to a field execution mechanism through an industrial Ethernet. The remote data acquisition control system belongs to the technical field of industrial automation control.

Description

Technical Field

[0001] This invention relates to the field of industrial automation control technology, specifically to a remote data acquisition and control system for a spray product filling production line. Background Technology

[0002] Spray product filling production lines involve the real-time acquisition and coordinated control of heterogeneous data from multiple sources, such as temperature, pressure, flow rate, and liquid level. Existing technologies are mostly based on data acquisition systems with fixed sampling periods and proportional-integral-derivative (PID) controllers. Fixed sampling period acquisition strategies are ill-suited to the varying rates of data change at different stages of the filling process. Higher sampling frequencies are required during the drastic parameter changes at the beginning and end of the filling process, while continuous high-frequency sampling during the stable filling phase generates a large amount of redundant data and increases network transmission burden. In nonlinear, time-varying, multivariable coupled systems, the control accuracy of PID controllers is easily affected by external disturbances, parameter tuning relies on manual experience, and the control law cannot be adaptively adjusted online when operating conditions change.

[0003] In multi-source sensor data fusion, existing solutions often set independent thresholds for each sensor's data to trigger alarms, or simply stitch the data together and input it into a single control model. The independent threshold approach severs the dynamic coupling relationship between parameters such as temperature, pressure, and flow rate during the filling process, easily leading to missed alarms when one parameter exceeds the limit while others remain within normal range, or misjudging small fluctuations in multiple parameters as abnormal. The simple stitching method fails to extract the deep interaction information between different sensor data in the feature space, resulting in an inaccurate representation of the production line status upon which subsequent control decisions are based, affecting the reliability of control commands.

[0004] In terms of optimizing control parameters, traditional gradient descent optimization algorithms rely on the continuous differentiability of the objective function. When faced with multiple extreme values ​​and nonlinear constraints in the filling process, they are prone to getting stuck in local optima and cannot find the combination of control parameters that makes the overall performance of the entire production line reach the global optimum, resulting in fluctuations in filling accuracy and high equipment energy consumption. Summary of the Invention

[0005] This invention provides a control system that can adaptively adjust the acquisition strategy based on the time-varying characteristics of data during the filling process of spray products and perform deep feature fusion of multi-source heterogeneous data. It enhances the global search capability of the parameter optimization algorithm through a chaotic mechanism, realizes online and accurate optimization of filling control parameters, and reliably sends control commands to the field actuators via industrial Ethernet.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a remote data acquisition and control system for a spray product filling production line. This system includes a data acquisition module, a feature fusion module, a target matching module, a parameter optimization module, and a command sending module. The data acquisition module initializes a particle swarm optimization population based on chaotic mapping to acquire multi-source heterogeneous data from the spray product filling production line. This method enables a more uniform initial distribution of sampling period parameters, avoiding data acquisition from falling into local optima, thereby improving the acquisition efficiency and representativeness of multi-source heterogeneous data. The feature fusion module performs feature fusion on the multi-source heterogeneous data using an adaptive fuzzy neural network to generate a state feature vector of the production line. This maps heterogeneous sensor information to a high-dimensional feature space, effectively suppressing single-source data noise and providing an accurate description of the production line's operating status for subsequent decision-making.

[0007] As a technical solution of this invention, when acquiring multi-source heterogeneous data based on chaotic mapping, Logistic chaotic mapping is used to generate initial positions and initial velocities for each particle in the particle swarm population. The initial position of each particle is mapped to the sampling period parameter of the corresponding data acquisition node, and the initial velocity of each particle is mapped to the adjustment step size of the sampling period parameter. According to these sampling period parameters, the data acquisition node is controlled to read the output values ​​of the temperature sensor, pressure sensor, flow meter, and level sensor respectively, thereby obtaining multi-source heterogeneous data. Preferably, when generating initial positions and initial velocities, the initial values ​​of the chaotic mapping are first randomly selected, and a chaotic sequence is iteratively generated according to the Logistic chaotic mapping formula; the chaotic sequence is linearly mapped to the upper and lower limits of the particle positions to obtain the initial position of each particle; the chaotic sequence is then subjected to a differential transformation and linearly mapped to the upper and lower limits of the particle velocities to obtain the initial velocity of each particle.

[0008] In this invention, when the adaptive fuzzy neural network performs feature fusion, temperature, pressure, flow rate, and liquid level values ​​from multi-source heterogeneous data are input into the first layer input nodes of the network. In the second layer, a Gaussian membership function is used to calculate the membership value corresponding to each input node. In the third layer, fuzzy rules are used to match the membership values ​​of all input nodes, with each fuzzy rule corresponding to a rule node, and the output of the rule node is the product of the input membership values. In the fourth layer, the outputs of all rule nodes are weighted and summed, with the weighting coefficients adaptively adjusted through network training. The result of the weighted summation is the state feature vector. As a preferred embodiment, the center and width parameters of the Gaussian membership function are obtained offline pre-trained using a backpropagation algorithm combined with historical operating data of the production line, enabling the fuzzy membership parameters to converge to a reasonable range in advance, thus accelerating the response speed of online fusion. Furthermore, after generating the state feature vector, the state feature vectors generated at multiple consecutive sampling times are stored in a circular buffer queue. After temporal alignment of the state feature vectors in the queue, the median of each dimension is calculated. The median vector is then used as the smoothed state feature vector for the current time step. This process can eliminate feature mutation spikes caused by occasional interference and improve the stability and reliability of the state representation.

[0009] The target matching module performs similarity matching between the state feature vector and preset filling process parameter templates to determine the control targets for the current production stage. Specifically, multiple filling process parameter templates are read from the database. Each template contains a standard state feature sequence and a corresponding set of control target parameters. A dynamic time warping algorithm is used to calculate the cumulative distance between the current state feature vector and each standard state feature sequence. The filling process parameter template corresponding to the standard state feature sequence with the smallest cumulative distance is selected as the matching template. Control targets are extracted from the matching templates, including the upper limit of target temperature, the lower limit of target temperature, the upper limit of target pressure, the lower limit of target pressure, the upper limit of target flow rate, and the lower limit of target flow rate. Preferably, the global path constraint window width in the dynamic time warping algorithm is set to twice the dimension of the state feature vector, which limits the offset range of the search path while ensuring matching accuracy and reducing computational complexity. Through this similarity matching mechanism, the system can accurately identify the current process stage of the production line, providing constraint boundaries that meet process requirements for parameter optimization.

[0010] The parameter optimization module uses a chaotic particle swarm optimization algorithm to iteratively optimize control parameters and generate remote control commands. In this module, the position vector of each particle is decoded into the proportional valve opening value, heater power value, and conveyor motor speed value. The fitness function value of each particle is calculated based on the deviation between the control target and the currently collected real-time temperature, pressure, and flow rate. This fitness function value is the weighted sum of squares of each deviation, thus comprehensively quantifying the control effect. In each iteration, a chaotic perturbation is used to update the particle velocity. This perturbation is generated by substituting the current particle position into a Logistic chaotic mapping. An adaptive inertia weight coefficient is introduced, which decreases linearly from its initial value to its final value with each iteration. The velocity is updated by multiplying the inertia weight coefficient by the original particle velocity and then adding the chaotic perturbation. The particle position is updated based on the updated velocity, and the fitness function value of the new position is calculated. The position of the globally optimal particle is recorded. A larger inertia weight in the early stages of iteration is beneficial for global search, while a smaller inertia weight in the later stages promotes local refinement. The addition of chaotic perturbation allows particles to escape local extrema, thereby quickly approaching the globally optimal combination of control parameters. As a preferred technical solution, after recording the position of the globally optimal particle, it is determined whether the current iteration count has reached the preset maximum iteration count. If not, the fitness function value of the globally optimal particle is compared with the fitness function value of the previous round. If the difference is less than the convergence threshold, the iteration ends early. If the maximum iteration count is reached or the early termination condition is met, the position of the current globally optimal particle is decoded to obtain the final proportional valve opening value, heater power value, and conveyor motor speed value. Then, according to the communication protocol used by the field actuator, the above control parameter values ​​are encapsulated into a remote control command frame containing a frame header, device address, instruction type, control parameter value, and frame check byte. This convergence judgment strategy avoids unnecessary iterative calculations and ensures the real-time nature of instruction generation.

[0011] The command sending module transmits remote control commands to field actuators via industrial Ethernet. This module employs a publish-subscribe model, establishing a message publishing node at the central control terminal. This node encapsulates the remote control commands into data packets according to the OPC unified architecture protocol. These data packets are then sent to the field industrial gateways via industrial Ethernet. The field industrial gateways run subscription nodes to receive the data packets. The field industrial gateways parse the data packets to obtain the various control values ​​from the remote control commands. An analog output module converts the proportional valve opening value into a voltage signal and outputs it to the proportional valve. A relay module converts the heater power value into a pulse-width modulation signal and outputs it to the heater. A digital output module converts the conveyor motor speed value into a frequency signal and outputs it to the frequency converter. The combination of the publish-subscribe model and the OPC unified architecture protocol enables cross-platform, highly reliable transmission of control commands in heterogeneous network environments. Preferably, the field industrial gateway has a built-in network interruption resume buffer. When a network connection interruption with the central control terminal is detected, the received remote control commands are cached locally and executed sequentially according to their timestamps, preventing the loss of control commands due to network interruptions and maintaining the continuous and stable operation of the production line.

[0012] As a further improvement of this invention, after sending the remote control command, the system receives the execution status code returned by the field actuator and the real-time feedback value from the sensors after the execution action. The execution status code includes a normal completion code and a fault alarm code. The execution status code and the real-time sensor feedback value are stored in the feedback buffer and compared with the current control target to calculate the feedback deviation vector. This feedback deviation vector is used as a correction term in the fitness function in the next round of iterative optimization. The weight coefficient of this correction term is dynamically adjusted according to the level of the fault alarm code. This method quantifies the execution consequences and feeds them back to the optimization algorithm. When a fault alarm occurs, the correction weight is increased, causing the optimization direction to quickly move away from dangerous operating conditions, forming a closed-loop control system with fault tolerance.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0014] The data acquisition module uses Logistic chaotic mapping to generate the initial position and initial velocity of each particle in the particle swarm population. The initial position of each particle is mapped to the sampling period parameter of the corresponding data acquisition node, and the initial velocity is mapped to the adjustment step size of the sampling period parameter. The ergodicity and pseudo-randomness of the chaotic mapping ensure that the particle swarm is evenly distributed in the sampling period value space, avoiding premature convergence in the subsequent parameter optimization process due to the initial population clustering in local areas. Compared with random initialization, this expands the initial search range, allowing the sampling period configuration scheme to cover a wider variety of sampling rhythm combinations. The data acquisition node reads the output values ​​of the temperature sensor, pressure sensor, flow meter, and level sensor according to the sampling period parameter corresponding to each particle. Different particles correspond to different sampling frequency configurations, and multiple sampling schemes run in parallel. This allows for the capture of sampling point distributions that better match the data change rate at different stages of filling, providing multi-source heterogeneous data with higher information density for subsequent feature fusion and control decisions.

[0015] The parameter optimization module decodes the position vector of each particle into the proportional valve opening value, heater power value, and conveyor motor speed value. It calculates the fitness function value based on the deviation between the control target and the current real-time temperature, pressure, and flow rate. In each iteration, a perturbation generated by the particle position through a Logistic chaotic mapping is introduced to update the particle velocity. An adaptive inertial weight coefficient, linearly decreasing from the initial value to the final value, is used. The inertial weight coefficient is multiplied by the original particle velocity and then the chaotic perturbation is added to complete the velocity update. The chaotic perturbation applies unpredictable small offsets during the particle velocity update process, allowing particles to still have the probability of escaping local optima in the later stages of the search. The adaptive inertial weight maintains a large value in the early stages of iteration to facilitate global exploration and gradually decreases in the later stages to enhance the local fine-grained search capability. The combination of these two approaches enables the algorithm to maintain population diversity within the multi-extreme nonlinear solution space of the filling process parameters, overcoming the shortcomings of traditional gradient optimization methods that rely on initial point selection and are easily captured by local extrema. The obtained proportional valve opening value, heater power value, and conveyor motor speed value are closer to the globally optimal control combination, and the dynamic fluctuations of temperature, pressure, and flow rate during the filling process are effectively suppressed. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a schematic diagram of a remote data acquisition and control system used in a spray product filling production line;

[0018] Figure 2This is a flowchart of particle swarm initialization and data acquisition based on chaotic mapping;

[0019] Figure 3 This is a flowchart of the target matching module;

[0020] Figure 4 This is a flowchart of parameter optimization and remote control command generation based on the chaotic particle swarm algorithm. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See Figure 1 This invention provides a remote data acquisition and control system for a spray product filling production line, comprising a data acquisition module, a feature fusion module, a target matching module, a parameter optimization module, and a command sending module. The data acquisition module initializes a particle swarm optimization population based on chaotic mapping and acquires multi-source heterogeneous data from the spray product filling production line. The feature fusion module performs feature fusion on the multi-source heterogeneous data using an adaptive fuzzy neural network to generate a state feature vector for the production line. The target matching module performs similarity matching between the state feature vector and a preset filling process parameter template to determine the control target for the current production stage. The parameter optimization module iteratively optimizes the control parameters using a chaotic particle swarm optimization algorithm to generate remote control commands. The command sending module sends the remote control commands to the field actuators via an industrial Ethernet network.

[0023] In specific implementation, please refer to Figure 2 The data acquisition module initializes the particle swarm population based on chaotic mapping and uses the initialized particle swarm to control the sampling behavior of the data acquisition nodes in order to collect multi-source heterogeneous data from the spray product filling production line. The process of initializing the particle swarm population includes three steps: chaotic generation of particle position and particle velocity, mapping of particle position and particle velocity to acquisition parameters, and driving the sensor to read data based on the acquisition parameters.

[0024] The initial position and initial velocity of the particles are generated using a Logistic chaotic mapping. The initial values ​​of the chaotic mapping are randomly selected and denoted as follows: , The value of is in the range (0,1) and not equal to 0.5, to avoid the chaotic sequence from entering a fixed point. The chaotic sequence is generated iteratively according to the Logistic chaotic mapping formula, which is expressed as:

[0025]

[0026] in: For chaos control parameters, A value of 4 causes the system to be in a completely chaotic state. For the first Chaotic values ​​generated by step-by-step iteration For the first Chaotic values ​​generated by step-by-step iteration For the iteration step index, The value of starts from 0 and continues until a chaotic sequence of the required length is generated. The length of the chaotic sequence is determined by the particle swarm size, which is set to include _____. If there are 100 particles, then it is necessary to generate 100 particles. A chaotic value serves as the basis for the initial position of the particle, and is obtained through difference transformation. A chaotic value serves as the basis for the initial velocity of the particle, therefore a total of [number] chaotic values ​​need to be generated. There are several chaotic values, where the first chaotic value is... Subsequent iterations generate to Population size Based on the real-time requirements of the data acquisition system and the pre-set computing resources, typically, The value is 30.

[0027] By linearly mapping the chaotic sequence to the upper and lower bounds of particle positions, the initial position of each particle is obtained. Taking the first... Taking a single particle as an example, its initial position The calculation is performed as follows: extract the chaotic value at the corresponding position in the chaotic sequence, and apply the linear mapping formula.

[0028]

[0029] in: For the chaotic sequence corresponding to the first The chaos value of an individual particle. This represents the upper limit of the particle's position. This represents the lower limit of the particle position. The upper and lower limits of the particle position are determined based on the configurable range of the sampling period of the data acquisition node. The sampling period parameter represents the time interval between two consecutive acquisitions of data from the same sensor by the data acquisition node. The upper limit of the sampling period parameter is set to 10 seconds, and the lower limit is set to 0.1 seconds. Therefore, the upper limit of the particle position... The value is 10.0, and the lower limit is... The value is set to 0.1. After performing a difference transformation on the chaotic sequence, it is linearly mapped to the range between the upper and lower limits of particle velocities to obtain the initial velocity of each particle. The difference transformation process is as follows: for the ... For each particle, take two adjacent chaotic values ​​from the chaotic sequence. and Calculate the difference between the two. , the difference As the mapping input, the difference sequence obtained from the difference transform is linearly mapped to the upper and lower limits of the particle velocity to obtain the initial velocity. The calculation method is as follows:

[0030]

[0031] in: This is the upper limit of particle velocity. This represents the lower limit of particle velocity. The upper and lower limits of particle velocity are determined based on the adjustment step size range of the sampling period parameter. The adjustment step size of the sampling period parameter represents the maximum allowable change each time the sampling period is updated. The upper limit of the adjustment step size is set to 1.0 second, and the lower limit is set to -1.0 second, meaning that the absolute value of the step size allowed for a single adjustment does not exceed 1.0 second. Therefore, the upper limit of particle velocity... The value is 1.0, and the lower limit is... The value is -1.0. After obtaining the initial position and initial velocity of each particle, the initial position of each particle is mapped to the sampling period parameter of the corresponding data acquisition node, and the initial velocity of each particle is mapped to the adjustment step size of the sampling period parameter. Specifically, for the _ Each particle, its sampling period parameter That is, equal to the initial position value of the particle. The unit is seconds; its sampling period adjustment step size That is, equal to the initial velocity value of the particle. The unit is seconds.

[0032] Data acquisition nodes are controlled to acquire data according to the sampling period parameters corresponding to each particle. For the first... The sampling strategy represented by each particle, and the data acquisition node using the sampling period parameter The system periodically reads the output values ​​of temperature sensors, pressure sensors, flow meters, and level sensors. The temperature sensor output is a digital value obtained by converting analog temperature signals from thermocouples or resistance thermometers installed in the filling machine's heating unit and storage tank into digital signals. The pressure sensor output is a digital value obtained by converting analog pressure signals from piezoresistive pressure transmitters installed in the inflation circuit and filling pipeline into digital signals. The flow meter output is a digital value of instantaneous flow rate obtained by counting and converting flow pulse signals from turbine flow meters or electromagnetic flow meters installed in the inflation pipeline and material pipeline into digital signals. The level sensor output is a digital value of level obtained by converting analog liquid level signals from ultrasonic level meters or differential pressure level transmitters installed in the storage tank into digital signals. In a single data acquisition operation, the data acquisition node sequentially reads the output values ​​of these four types of sensors, forming a four-dimensional data tuple, which represents multi-source heterogeneous data. After each sampling cycle, the data acquisition node repeats the above reading operation, thereby continuously collecting temperature, pressure, flow rate and liquid level data of the spray product filling production line, and sending the multi-source heterogeneous data at each moment to the feature fusion module.

[0033] It is understandable that after the specific particle swarm optimization process starts, the sampling period parameter of the data acquisition node is not fixed, but dynamically adjusted as the particle position is updated during the iterative optimization process. The step size of the sampling period adjustment then acts as a constraint when the particle velocity is updated. In the initial stage, the diverse initial positions and initial velocities generated by the Logistic chaotic mapping give the particle swarm population ergodicity and a uniformly distributed initial sampling period parameter, which helps to avoid the search process getting trapped in local optima and improves the comprehensiveness of multi-source heterogeneous data acquisition.

[0034] In practical implementation, the feature fusion module receives multi-source heterogeneous data output by the data acquisition module, performs feature fusion on the multi-source heterogeneous data through an adaptive fuzzy neural network, generates the state feature vector of the production line, and performs time-series smoothing on the state feature vector at multiple consecutive sampling times, outputting the smoothed state feature vector.

[0035] The adaptive fuzzy neural network employs a four-layer feedforward structure. The first layer is the input layer, containing four input nodes. These nodes correspond to the temperature, pressure, flow rate, and liquid level values ​​from the multi-source heterogeneous data. The temperature value is input to the first input node, the pressure value to the second, the flow rate to the third, and the liquid level to the fourth. Before entering the adaptive fuzzy neural network, the data input to each node undergoes normalization. This normalization maps the temperature, pressure, flow rate, and liquid level values ​​to the [0,1] interval. The lower and upper scaling limits used in the normalization process correspond to the minimum and maximum values ​​for each sensor range, respectively.

[0036] The second layer of the adaptive fuzzy neural network is the fuzzification layer. In this layer, a Gaussian membership function is used to calculate the membership value for each input node. The second layer contains multiple sets of membership function nodes, with each input node corresponding to one set. The number of nodes in each set is determined based on the precision of the input space partitioning covered by the fuzzy rules. (The last sentence appears to be incomplete and possibly refers to a different layer.) Taking the nth input node as an example, the nth Each input node corresponds to Membership function nodes The value is 5. The value of is determined by dividing the input space into five fuzzy subsets, corresponding to the five semantic labels: negative large, negative small, zero, positive small, and positive large. The first input node corresponding to the first input node The output of each membership function node is calculated using the Gaussian membership function, the expression of which is:

[0037]

[0038] in: For the first The normalized input values ​​received by each input node The value range is [0,1]; For the first The first input node The central parameter of a Gaussian membership function The initial value is obtained by uniformly dividing the input space [0,1] into five intervals and taking the midpoint of each interval; For the first The first input node The width parameter of a Gaussian membership function. The initial value is set to 0.1; For input values For the first The first input node The membership degree value of a node using the membership function. The central parameter of the Gaussian membership function. and width parameter It is obtained through offline pre-training using the backpropagation algorithm combined with historical operating data of the production line.

[0039] The specific process of offline pre-training is as follows: Historical operating data from multiple batches of continuous operation on the production line under normal filling conditions is collected, including temperature, pressure, flow rate, liquid level, and ideal state feature vectors labeled by process engineers according to the production stage at each sampling moment. The temperature, pressure, flow rate, and liquid level values ​​are normalized and used as input to the adaptive fuzzy neural network. The labeled ideal state feature vectors are used as the target output of the adaptive fuzzy neural network, and a mean squared error loss function is constructed. The mean squared error loss function represents the deviation between the actual output state feature vector of the adaptive fuzzy neural network and the labeled ideal state feature vector. Gradient descent is used to train the central parameters. and width parameter The update process involves setting the learning rate of gradient descent to 0.01. In each iteration, the partial derivatives of the loss function with respect to the center and width parameters are calculated. The values ​​of the center and width parameters are adjusted according to the negative direction of the partial derivatives. This iteration is repeated until the loss function converges to below a preset convergence threshold, which is set to 0.001. After training, the obtained Gaussian membership function center and width parameters are stored in the parameter memory of the adaptive fuzzy neural network for direct loading and use during the online inference phase.

[0040] The third layer of the adaptive fuzzy neural network is the rule layer. In this layer, fuzzy rules are applied to the membership values ​​of all input nodes, with each fuzzy rule corresponding to a rule node. The fuzzy rules are generated by matching the membership values ​​of each input node with fuzzy rules. Each membership function is arranged in a full combination permutation, and each combination constitutes a fuzzy rule. All combinations generate a total of [number missing] fuzzy rules. There are 625 fuzzy rules, corresponding to a total of 5×5×5×5 rules. The output of each rule node is the product of its input membership values. The membership function, the second input node's first... The membership function, the third input node's... The membership function and the fourth input node's first... The output value of the rule node corresponding to the fuzzy rule generated by the combination of membership functions. The calculation method is as follows:

[0041]

[0042] in: The normalized temperature value received by the first input node relative to the first input node's... The membership degree values ​​of a membership function. The normalized pressure value received by the second input node relative to the second input node's first... The membership degree values ​​of a membership function. The normalized value of the traffic received by the third input node relative to the third input node's... The membership degree values ​​of a membership function. The normalized liquid level value received by the fourth input node is relative to the fourth input node's... The membership degree values ​​of each membership function.

[0043] The fourth layer of the adaptive fuzzy neural network is the defuzzification layer. In this layer, the outputs of all rule nodes are weighted and summed. The weighting coefficients are adaptively adjusted during network training, and the result of the weighted sum is the state feature vector. The state feature vector is... dimensional vector, The value of is set to 4, which is the th value of the state feature vector. Dimensional components are denoted as , The value range of is 1, 2, 3, 4. Dimensional components The calculation method is as follows:

[0044]

[0045] in: For the first Dimensional output components and fuzzy rules The corresponding weighting coefficients are initially generated by random sampling from a normal distribution with a mean of 0 and a variance of 0.01. Weighting coefficients During offline pre-training, the center parameters of the Gaussian membership function are obtained through backpropagation algorithm. and width parameter Synchronous updates are performed. During backpropagation, the partial derivatives of the mean squared error loss function with respect to the weighted coefficients are calculated, and the value of each weighted coefficient is updated using the gradient descent method. The learning rate is consistent with the update learning rates of the center and width parameters, both set to 0.01. After training, the obtained weighted coefficients are stored in the parameter memory of the adaptive fuzzy neural network for direct loading and use during the online inference phase.

[0046] After generating the state feature vectors, the state feature vectors generated at multiple consecutive sampling times are stored in a circular buffer queue. The circular buffer queue is a fixed-length queue with a first-in, first-out (FIFO) structure, and its length is set to 5 sampling times. This queue length is designed to suppress random fluctuations while avoiding excessive lag. At each current sampling time, the state feature vector generated by the adaptive fuzzy neural network is pushed to the tail of the circular buffer queue. Simultaneously, the state feature vector with the longest remaining time at the head of the queue is automatically removed, ensuring the circular buffer queue always contains the state feature vectors from the most recent 5 sampling times. After temporal alignment of the state feature vectors in the circular buffer queue, the median of each dimension is calculated, and the median vector is used as the smoothed state feature vector output for the current time. Temporal alignment refers to sorting the 5 state feature vectors in the circular buffer queue according to their sampling timestamps from earliest to latest. For the sorted 5 state feature vectors, the median of each dimension is calculated. Five values ​​are extracted from each dimension. These five values ​​are then arranged in ascending order, and the third value in the sorted sequence is taken as the first value. The median of each dimension is calculated. The median calculation is performed on each of the four dimensions of the state feature vector. The resulting four-dimensional vector, composed of the medians of the four dimensions, is the smoothed state feature vector at the current sampling time. This smoothed state feature vector is then sent to the target matching module for subsequent similarity matching and target determination.

[0047] In specific implementation, please refer to Figure 3 The target matching module receives the smoothed state feature vector output by the feature fusion module, and performs similarity matching between the state feature vector and the preset filling process parameter template to determine the control target of the current production stage.

[0048] Multiple filling process parameter templates are read from the database. The database is a relational database or time-series database built on non-volatile storage media, pre-stored with filling process parameter templates corresponding to various typical production stages of the spray product filling production line. Each filling process parameter template contains a standard state characteristic sequence and a corresponding set of control target parameters. The standard state characteristic sequence is a sequence of state characteristic vectors that, as confirmed by process engineers, represent the ideal operating state of a specific production stage, arranged in chronological order from historical production batches. Each element in the standard state characteristic sequence is a four-dimensional state characteristic vector, with each dimension corresponding to a temperature state characteristic value, a pressure state characteristic value, a flow rate state characteristic value, and a liquid level state characteristic value, respectively. The set of control target parameters includes the target upper temperature limit, target lower temperature limit, target upper pressure limit, target lower pressure limit, target upper flow rate limit, and target lower flow rate limit. The upper and lower limits of the target temperature in the set of control target parameters are determined by process engineers based on the allowable temperature range for the filling quality of the spray product at this production stage. The upper and lower limits of the target pressure are determined by process engineers based on the allowable fluctuation range of the filling inflation pressure and the liquid delivery pressure. The upper and lower limits of the target flow rate are determined by process engineers based on the allowable range of the filling rate and the inflation flow rate. Upon system startup or during periodic refresh, the target matching module reads all filling process parameter templates from the database and loads them into the template cache in memory. Each entry in the template cache stores a template identifier, a standard state characteristic sequence, and the set of control target parameters for a filling process parameter template.

[0049] The dynamic time warping algorithm is used to calculate the cumulative distance between the current state feature vector and each standard state feature sequence. The current state feature vector refers to the smoothed state feature vector output by the feature fusion module at the current sampling time, which is a four-dimensional vector, denoted as . ,in The temperature state characteristic value at the current moment. The current pressure state characteristic value. This represents the current flow state characteristic value. This represents the current liquid level status characteristic value. For the th in the template buffer... A filling process parameter template, whose standard state characteristic sequence is as follows:

[0050]

[0051] in: For the first The length of the standard state characteristic sequence of a filling process parameter template For the first In the standard state characteristic sequence of the filling process parameter template, the first... The standard state feature vector at each time step It is a four-dimensional vector. ,in For the first The first template The standard temperature characteristic value at each moment, For the first The first template The standard pressure characteristic value at each moment, For the first The first template Standard flow characteristic value at a given time point For the first The first template The standard liquid level characteristic value at each moment. Calculate the state feature vector at the current moment. With the Standard state characteristic sequence of a filling process parameter template When calculating the cumulative distance between them, construct a The cumulative distance matrix. The first cumulative distance matrix. Line 1 Column element representation and Under a certain correspondence, the [number]th [item] is reached. The position of the first standard sequence and the first The cumulative distance at each query sequence position. The query sequence only contains the state feature vector at the current time, therefore the query sequence length is fixed at 1, and the lengths of the two sequences to be matched are respectively... And 1. In the dynamic time warping algorithm, the global path constraint window width is set to twice the dimension of the state feature vector. The dimension of the state feature vector is 4, so the global path constraint window width is set to 8. The global path constraint window width determines the maximum allowable range of index offsets between two sequences during alignment, when the standard state feature sequence length... When the difference between the template and the query sequence length 1 exceeds the window width 8, the filling process parameter template is determined to be mismatched with the current state feature vector, the cumulative distance is directly set to positive infinity, and it will no longer participate in the comparison of the minimum cumulative distance.

[0052] For a filling process parameter template that satisfies the window width constraint, start from the first standard state feature vector of the standard state feature sequence. Begin, calculate and The local distance between them is measured using Euclidean distance, and is calculated as follows: Traverse all of the standard state feature sequences Each standard state feature vector is calculated as... The Euclidean distance between them is obtained. The local distance value. The dynamic time warping algorithm recursively calculates the cumulative distance through dynamic programming, in order to... The Middle Each standard state feature vector and Based on the local distance between them, combined with the first The cumulative distance up to the th standard state feature vector is recursively obtained to obtain the th th . The cumulative distance of the steps. The recursive process is then transformed from... Execute to The cumulative distance obtained is the state feature vector at the current moment. With the Standard state characteristic sequence of a filling process parameter template The cumulative distance between them. Perform the above cumulative distance calculation on all filling process parameter templates to obtain the cumulative distance value corresponding to each filling process parameter template.

[0053] The filling process parameter template corresponding to the standard state feature sequence with the smallest cumulative distance is selected as the matching template. The cumulative distance values ​​corresponding to all filling process parameter templates are sorted in ascending order, and the template identifier of the filling process parameter template with the smallest cumulative distance value is taken. The set of control target parameters for this filling process parameter template is read from the template buffer.

[0054] The control targets are extracted from the matching template. These targets include the upper limit of target temperature, the lower limit of target temperature, the upper limit of target pressure, the lower limit of target pressure, the upper limit of target flow rate, and the lower limit of target flow rate. The upper and lower limits of target temperature are combined to form a temperature control interval, the upper and lower limits of target pressure are combined to form a pressure control interval, and the upper and lower limits of target flow rate are combined to form a flow control interval. These control targets are then passed to the parameter optimization module as a reference benchmark for calculating the fitness function.

[0055] In specific implementation, please refer to Figure 4 The parameter optimization module receives the control target output by the target matching module, iteratively optimizes the control parameters using a chaotic particle swarm optimization algorithm, and generates remote control commands. First, the chaotic particle swarm optimization algorithm population is initialized. The particle swarm population size is set to [value missing]. , The value is 30. The value of the parameter is determined by balancing the time consumed in iterative computation while ensuring sufficient coverage of the search space. Each particle represents a set of candidate control parameters, including the proportional valve opening value, heater power value, and conveyor motor speed value. Therefore, the position vector of each particle is a three-dimensional vector. The position vector of each particle is decoded into the proportional valve opening value, heater power value, and conveyor motor speed value. The decoding method is direct mapping, that is, the first dimension of the particle position vector directly corresponds to the proportional valve opening value, in percentage, with the upper limit of the search range of the proportional valve opening value set to 100% and the lower limit set to 0%; the second dimension of the particle position vector directly corresponds to the heater power value, in kilowatts, with the upper limit of the search range of the heater power value set to the rated power of the heater and the lower limit set to 0 kilowatts; the third dimension of the particle position vector directly corresponds to the conveyor motor speed value, in revolutions per minute, with the upper limit of the search range of the conveyor motor speed value set to the rated speed of the conveyor motor and the lower limit set to 0 revolutions per minute. The initial positions of all particles in the particle swarm are generated using a Logistic chaotic mapping. The specific generation process is as follows: a random initial value for the chaotic mapping, between 0 and 1 and not equal to 0.5, is selected, and the positions are generated according to the Logistic chaotic mapping formula. A chaotic sequence is iteratively generated and linearly mapped to the search ranges of the proportional valve opening value, the heater power value, and the conveyor motor speed value, respectively, to obtain the initial position of each particle. The initial velocity of each particle is generated using the same chaotic mapping differential transformation method, with the upper and lower limits of the particle velocity set to positive and negative values ​​of one-tenth of the corresponding search range span, respectively.

[0056] After iterative optimization begins, in each iteration, the fitness function value of each particle is calculated based on the deviation between the control target and the currently collected real-time temperature, real-time pressure, and real-time flow rate. The currently collected real-time temperature is the latest temperature value read from the temperature sensor, the real-time pressure is the latest pressure value read from the pressure sensor, and the real-time flow rate is the latest flow rate value read from the flow meter. The control target includes the upper limit of the target temperature, the lower limit of the target temperature, the upper limit of the target pressure, the lower limit of the target pressure, the upper limit of the target flow rate, and the lower limit of the target flow rate. For the... Each particle, after having its position vector decoded to obtain the proportional valve opening value, heater power value, and conveyor motor speed value, applies these values ​​to the field actuator. This changes the operating status of the spray product filling production line. The data acquisition module continuously collects real-time data, using the collected real-time temperature, pressure, and flow rate as the response values ​​under the corresponding control parameters for that particle. The fitness function value is the weighted sum of squares of the deviations, and its expression is:

[0057]

[0058] in: For the first The fitness function value of each particle. The smaller the value, the closer the control parameter corresponding to the particle can make the real-time state of the production line approach the control target. This is the temperature deviation weighting coefficient. The value is 0.5. The value is determined based on the balance between the influence of temperature on the filling quality of spray products and pressure and flow rate. This is the pressure deviation weighting coefficient. The value is 0.3. The value is determined based on the fact that the impact of pressure fluctuations on filling consistency is slightly lower than that of temperature. The flow deviation weighting coefficient is used. The value is 0.2. The value is based on the fact that the flow deviation can be indirectly compensated through liquid level linkage in closed-loop regulation; This is the temperature deviation term, calculated as follows: when the real-time temperature value is between the lower limit of the target temperature and the upper limit of the target temperature... The value is 0 when the real-time temperature is lower than the lower limit of the target temperature. The value is taken as the lower limit of the target temperature minus the real-time temperature value. When the real-time temperature value is higher than the upper limit of the target temperature value, The value is the real-time temperature value minus the target temperature upper limit. This is the pressure deviation term. The calculation method for the pressure deviation term is the same as that for the temperature deviation term. It is obtained by comparing the real-time pressure value with the target lower limit and the target upper limit. The flow rate deviation term is calculated in the same way as the temperature deviation term, by comparing the real-time flow rate value with the target flow rate lower limit and the target flow rate upper limit. In each iteration, chaotic perturbation is used to update the particle velocity. Chaotic perturbation generates the perturbation quantity by substituting the current particle position into the Logistic chaotic mapping. For the... For each particle, the first dimension of its position vector at the current iteration step is normalized to the (0,1) interval and used as the input value for the Logistic chaotic mapping. A chaotic value is calculated, where Indicates the first The normalized value of the first dimension component of the particle position vector. The calculated chaotic value is linearly mapped to the interval [-0.1, 0.1] to obtain a three-dimensional chaotic perturbation vector. Each dimension of the chaotic perturbation vector uses the same mapped chaotic value.

[0059] An adaptive inertia weight coefficient is introduced, which decreases linearly from its initial value to its final value with the number of iterations. Let the current iteration number be... The preset maximum number of iterations is , The value is 100. The value is determined based on the algorithm's ability to converge stably within this number of iterations under typical operating conditions. (Inertia weight coefficient) The calculation method is as follows:

[0060]

[0061] in: This is the initial value for the inertia weighting coefficient. The value is 0.9. The value is chosen to give the algorithm a strong global search capability in the early stages of iteration; This is the final value of the inertia weighting coefficient. The value is 0.4. The chosen value is based on enabling the algorithm to possess strong local fine-grained search capabilities in the later stages of iteration. The inertia weight coefficient is multiplied by the original particle velocity, and then the chaotic perturbation update rate is added. The updated velocity vector of each particle The calculation method is as follows:

[0062]

[0063] in: For the first The velocity vector of each particle before the update For the first The current position vector of each particle. For the first The individual historical optimal position vector of each particle. This represents the position vector of the globally optimal particle. and As a learning factor, The value is 2.0. The value is 2.0. and The random numbers are uniformly distributed in the interval [0,1] and are regenerated randomly in each iteration. Let be the chaotic perturbation vector. Update the particle position based on the updated velocity, the th... The updated position vector of each particle The calculation method is as follows After the position is updated, check if each dimension of the new position vector exceeds the corresponding upper or lower bound of the search range. If it exceeds the upper bound, set that dimension component to the upper bound of the search range; if it exceeds the lower bound, set that dimension component to the lower bound of the search range. Calculate the fitness function value of the new position, using the same method as the fitness function value calculation described earlier. Compare the fitness function value of the new position with the... The individual historical best fitness function value of each particle is compared. If the fitness function value at the new position is smaller, then the particle is moved to the next position. The individual historical best position vector of each particle is updated to the new position vector. Then, all particles are traversed. The individual historical best fitness function value of each particle is used as the position vector of the globally best particle in this iteration. The position vector of the particle with the smallest individual historical best fitness function value is then recorded. Finally, it is determined whether the current iteration count has reached the preset maximum iteration count. If the maximum number of iterations has not been reached, the fitness function value of the globally optimal particle is compared with the globally optimal fitness function value of the previous round. If the difference is less than the convergence threshold, the iteration ends early. The convergence threshold is set to 0.001, and its value is based on the assumption that fluctuations in the fitness function value below this level have a negligible impact on the actual adjustment of the control parameters. The difference is calculated by subtracting the globally optimal fitness function value of the current round from the globally optimal fitness function value of the previous round and taking the absolute value.

[0064] If the maximum number of iterations is reached or the early termination condition is met, the position vector of the current globally optimal particle is decoded to obtain the final proportional valve opening value, heater power value, and conveyor motor speed value. The decoding method is as follows: the first dimension of the position vector of the globally optimal particle is the final proportional valve opening value, the second dimension is the final heater power value, and the third dimension is the final conveyor motor speed value.

[0065] According to the communication protocol used by the field actuators, the proportional valve opening value, heater power value, and conveyor motor speed value are encapsulated into a remote control command frame. The remote control command frame includes a frame header, device address, command type, control parameter value, and frame check byte. The frame header is a fixed one-byte value, 0xA5, used to identify the start of the command frame. The device address occupies one byte and is used to distinguish different field actuators; the device address for the proportional valve is set to 0x01, the device address for the heater is set to 0x02, and the device address for the conveyor motor inverter is set to 0x03. The command type occupies one byte, fixed at 0x10, indicating that this command frame is a control parameter write command. The control parameter value occupies two bytes. For the instruction frame corresponding to the proportional valve device address, the control parameter value is stored as an unsigned integer, representing the percentage value of the proportional valve opening multiplied by 10. For the instruction frame corresponding to the heater device address, the control parameter value is stored as an unsigned integer, representing the kilowatt value of the heater power multiplied by 100. For the instruction frame corresponding to the conveyor motor inverter device address, the control parameter value is stored as an unsigned integer, representing the conveyor motor speed in minutes. The frame check byte occupies one byte and is calculated using a cyclic redundancy check algorithm on all bytes of the four fields: frame header, device address, instruction type, and control parameter value. After the remote control instruction frame is generated, the three remote control instruction frames corresponding to the three device addresses are sent sequentially to the instruction sending module.

[0066] In practice, the instruction sending module receives the remote control instruction frame generated by the parameter optimization module and sends the remote control instruction to the field actuator via industrial Ethernet.

[0067] A publish-subscribe model is adopted, establishing message publishing nodes at the central control terminal. The central control terminal is an industrial computer or server running the core program of the data acquisition and control system. The message publishing node is a logical communication entity created on the central control terminal based on message middleware software. The message publishing node encapsulates remote control commands into data packets according to the OPC unified architecture protocol. The OPC unified architecture protocol adopts an information model combining client-server and publish-subscribe. The message publishing node, as a publisher, registers a topic node in the address space. The naming format of the topic node is " / SprayFilling / ControlCommands / ActuatorType", where ActuatorType corresponds to "ProportionalValve", "Heater", and "ConveyorMotor" respectively according to the device address in the remote control command frame. The data packet encapsulation process is as follows: the frame header, device address, command type, control parameter value, and frame check byte of the remote control command frame are used as the payload, plus the dataset message header of the OPC unified architecture protocol. The dataset message header contains a timestamp, publisher identifier, and dataset name, forming a complete data packet.

[0068] Data packets are sent to the field industrial gateway via industrial Ethernet. Industrial Ethernet uses industrial communication protocol stacks such as EtherNet / IP or Profinet, with shielded twisted-pair cable or fiber optic cable as the physical link, and a transmission rate of 100Mbps or 1Gbps. The industrial Ethernet interface card at the central control end converts the data packets into physical layer signals for transmission.

[0069] The field industrial gateway operates by subscribing nodes to receive data packets. The field industrial gateway is an embedded communication device deployed on the production line, equipped with at least one industrial Ethernet interface and multiple fieldbus interfaces. Upon startup, the field industrial gateway sends a subscription request to the message publishing node at the central control end, specifying the topic nodes to be subscribed to. After receiving the subscription request, the message publishing node establishes a publish-subscribe relationship and pushes subsequently generated remote control command data packets to the field industrial gateway. The subscribing nodes of the field industrial gateway continuously listen to the network port, receiving data packets pushed by the publishing node.

[0070] The on-site industrial gateway parses data packets to obtain the proportional valve opening value, heater power value, and conveyor motor speed value from the remote control commands. The parsing process is as follows: extract the dataset message header of the OPC unified architecture protocol from the data packet, verify that the publisher identifier matches the publisher identifier in the subscription request, strip the dataset message header, extract the payload, parse it byte by byte according to the format of the remote control command frame, and map the control parameter values ​​to the proportional valve opening value, heater power value, or conveyor motor speed value according to the value of the device address field.

[0071] The field industrial gateway converts the proportional valve opening value into a voltage signal and outputs it to the proportional valve via an analog output module. The analog output module is an analog output board that extends the field industrial gateway, with a resolution of at least 12 bits and an output voltage range of 0 to 10 volts. A 0% proportional valve opening value corresponds to an output of 0 volts, and 100% corresponds to an output of 10 volts. The conversion formula is as follows:

[0072]

[0073] in: This is the output voltage value, in volts. This represents the proportional valve opening value, expressed as a percentage. The analog output module outputs the converted voltage signal to the analog control input of the proportional valve. The field industrial gateway converts the heater power value into a pulse-width modulation (PWM) signal and outputs it to the heater via a relay module. The relay module is a solid-state relay or electromagnetic relay assembly with a maximum switching frequency of at least 1Hz. The period of the PWM signal is set to 2 seconds. A 100% heater power value corresponds to a 100% duty cycle, and a 50% power value corresponds to a 50% duty cycle. The high-level duration of the PWM signal is:

[0074]

[0075] in: The duration of the high-level signal, in seconds. This represents the heater power value as a percentage. The relay module's contact output is connected to the AC contactor coil or the thyristor trigger terminal of the heater's power supply circuit, controlling the heater's average power via a pulse width modulation signal.

[0076] The on-site industrial gateway converts the conveyor motor speed into a frequency signal and outputs it to the frequency converter via a digital output module. The digital output module is a pulse output channel with a maximum output frequency of at least 10kHz. The linear mapping between the conveyor motor speed and the output frequency signal is as follows: 0 rpm corresponds to 0Hz output frequency, and the rated speed of the conveyor motor corresponds to the upper limit of the output frequency. The upper limit of the output frequency is set to the highest input frequency reference value of the frequency converter, typically 50Hz or 60Hz. The digital output module connects the frequency signal to the analog or pulse input terminals of the frequency converter via a twisted-pair cable.

[0077] The field industrial gateway has a built-in network outage resume buffer, which is a segment of storage space in the gateway's non-volatile memory, configured to cache the most recent 1000 remote control commands. When the field industrial gateway detects a network connection interruption with the central control unit, the interruption is detected by failing to receive a response from the central control unit for three consecutive heartbeat packet transmission cycles (each heartbeat packet transmission cycle is set to 1 second). After detecting a network interruption, the field industrial gateway caches the received remote control commands in the network outage resume buffer, with each cached command accompanied by a timestamp. The field industrial gateway retrieves the remote control commands from the network outage resume buffer one by one according to the timestamps, parses and executes them, ensuring that commands are not lost during the network interruption and are executed in chronological order. When the network connection is restored, the field industrial gateway continues to receive real-time remote control commands from the message publishing node, while simultaneously processing any unexecuted cached commands in the network outage resume buffer, until all cached commands have been executed, at which point it switches to a fully real-time command execution mode.

[0078] After sending the remote control command, the command sending module receives the execution status code returned by the field actuator and the real-time sensor feedback values ​​after the execution action. The execution status code is transmitted back to the central control terminal by the field actuator through the field industrial gateway. The execution status code includes a normal completion code and a fault alarm code. The normal completion code is defined as 0x00, and the fault alarm code is defined as a series of non-zero values ​​according to different fault types. For example, the proportional valve jamming fault alarm code is 0x11, the heater over-temperature fault alarm code is 0x12, and the conveyor motor overload fault alarm code is 0x13. The real-time sensor feedback values ​​are the latest output values ​​of the temperature sensor, pressure sensor, flow meter, and level sensor collected by the data acquisition module after the execution action.

[0079] The execution status code and real-time sensor feedback values ​​are stored in the feedback buffer. The feedback buffer is a circular buffer allocated in the central control unit's memory, storing feedback data from the most recent 10 sampling periods. Each storage unit contains the execution status code, real-time temperature feedback value, real-time pressure feedback value, real-time flow rate feedback value, and real-time liquid level feedback value. The latest real-time sensor feedback value in the feedback buffer is compared with the current control target to calculate the feedback deviation vector. The feedback deviation vector is a four-dimensional vector. ,in This refers to the temperature feedback deviation component, which is calculated as follows: when the real-time temperature feedback value is between the lower limit of the target temperature and the upper limit of the target temperature, ... The value is 0; otherwise, the absolute value of the difference between the real-time temperature feedback value and the nearest target temperature boundary value is taken. This is the pressure feedback deviation component, and its calculation method is the same as that of the temperature feedback deviation component. This is the flow feedback deviation component, calculated in the same way as the temperature feedback deviation component. The liquid level feedback deviation component is the difference between the real-time liquid level feedback value and the liquid level set value. The liquid level set value is determined based on the liquid level standard value implicit in the filling process parameter template.

[0080] The feedback bias vector is used as a correction term for the fitness function in the next round of iterative optimization. The weight coefficient of the correction term is dynamically adjusted according to the level of the fault alarm code. The expression for the corrected fitness function is:

[0081]

[0082] in: For the first The fitness function value after correction for each particle; For the first The original fitness function value of each particle is calculated based on temperature, pressure, and flow rate deviations; This is the temperature feedback deviation component; This refers to the pressure feedback deviation component. For flow feedback deviation component; This is the liquid level feedback deviation component; These are the weighting coefficients for the correction term. The value is dynamically determined based on the fault alarm code level in the execution status code: when the execution status code is a normal completion code 0x00... The value is 0.1; when the execution status code is a general fault alarm code (alarm code value range 0x10 to 0x1F), The value is 0.5; when the execution status code is a critical fault alarm code (alarm code value range 0x20 to 0xFF), The value is 1.0. The classification of fault alarm codes is based on the degree of impact of the fault on the filling process quality and equipment safety. General fault alarm codes correspond to fault types that can be automatically recovered from or have a minor impact on quality. Critical fault alarm codes correspond to fault types that require immediate adjustment of the control strategy or may cause batch defects in the product. (Modified fitness function value) In the next iteration of the chaotic particle swarm optimization algorithm in the parameter optimization module, the original fitness function value is replaced. Participating in particle evaluation and global optimal particle update enables the control command generation process to perform closed-loop correction based on actual execution feedback.

[0083] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A remote data acquisition and control system for a spray product filling production line, characterized in that, include: The data acquisition module initializes the particle swarm population based on chaotic mapping and collects multi-source heterogeneous data from the spray product filling production line. The feature fusion module performs feature fusion on the multi-source heterogeneous data through an adaptive fuzzy neural network to generate a state feature vector of the production line. The target matching module performs similarity matching between the state feature vector and the preset filling process parameter template to determine the control target for the current production stage. The parameter optimization module uses a chaotic particle swarm optimization algorithm to iteratively optimize the control parameters and generate remote control commands. The command sending module sends the remote control commands to the field actuators via industrial Ethernet.

2. The remote data acquisition and control system for a spray product filling production line according to claim 1, characterized in that, The step of initializing the particle swarm population based on chaotic mapping and collecting multi-source heterogeneous data from the spray product filling production line specifically includes: The initial position and initial velocity of each particle are generated using a Logistic chaotic mapping. The initial position of each particle is mapped to the sampling period parameter of the corresponding data acquisition node, and the initial velocity of each particle is mapped to the adjustment step size of the sampling period parameter. According to the sampling period parameters, the data acquisition nodes are controlled to read the output values ​​of the temperature sensor, pressure sensor, flow meter and liquid level sensor respectively to obtain multi-source heterogeneous data.

3. The remote data acquisition and control system for a spray product filling production line according to claim 2, characterized in that, The step of generating the initial position and initial velocity of each particle using Logistic chaotic mapping specifically includes: Randomly select initial values ​​for the chaotic map and iteratively generate a chaotic sequence according to the Logistic chaotic map formula; The initial position of each particle is obtained by linearly mapping the chaotic sequence to the upper and lower limits of the particle position. After performing a differential transformation on the chaotic sequence, it is linearly mapped to the upper and lower limits of particle velocity to obtain the initial velocity of each particle.

4. The remote data acquisition and control system for a spray product filling production line according to claim 1, characterized in that, The step of fusing features from the multi-source heterogeneous data using an adaptive fuzzy neural network to generate a state feature vector for the production line specifically includes: The temperature, pressure, flow rate, and liquid level values ​​from the multi-source heterogeneous data are respectively input into the first layer input node of the adaptive fuzzy neural network; In the second layer, the Gaussian membership function is used to calculate the membership value of each input node. In the third layer, fuzzy rule matching is performed on the membership values ​​of all input nodes. Each fuzzy rule corresponds to a rule node, and the output of the rule node is the product of the input membership values. In the fourth layer, the outputs of all rule nodes are summed with weights. The weighting coefficients are adaptively adjusted through network training, and the result of the weighted sum is the state feature vector.

5. The remote data acquisition and control system for a spray product filling production line according to claim 4, characterized in that, The center and width parameters of the Gaussian membership function are obtained through offline pre-training using the backpropagation algorithm combined with historical operating data of the production line.

6. The remote data acquisition and control system for a spray product filling production line according to claim 1, characterized in that, The steps for determining the control target of the current production stage by performing similarity matching between the state feature vector and the preset filling process parameter template specifically include: Multiple filling process parameter templates are read from the database. Each filling process parameter template contains a standard state characteristic sequence and a corresponding set of control target parameters. The dynamic time warping algorithm is used to calculate the cumulative distance between the current state feature vector and each standard state feature sequence; The filling process parameter template corresponding to the standard state feature sequence with the smallest cumulative distance is selected as the matching template; The control targets are extracted from the matching template. The control targets include the upper limit of the target temperature, the lower limit of the target temperature, the upper limit of the target pressure, the lower limit of the target pressure, the upper limit of the target flow rate, and the lower limit of the target flow rate.

7. The remote data acquisition and control system for a spray product filling production line according to claim 6, characterized in that, In the dynamic time warping algorithm, the global path constraint window width is set to twice the dimension of the state feature vector.

8. The remote data acquisition and control system for a spray product filling production line according to claim 1, characterized in that, The step of iteratively optimizing the control parameters using the chaotic particle swarm optimization algorithm to generate remote control commands specifically includes: The position vector of each particle is decoded into the proportional valve opening value, the heater power value, and the conveyor motor speed value; The fitness function value of each particle is calculated based on the deviation between the control target and the currently collected real-time temperature, real-time pressure, and real-time flow rate. The fitness function value is the weighted sum of squares of the deviations. In each iteration, the velocity of the particle is updated using a chaotic perturbation, which is generated by substituting the current position of the particle into a Logistic chaotic map. An adaptive inertia weight coefficient is introduced. The inertia weight coefficient decreases linearly from the initial value to the final value with the number of iterations. The inertia weight coefficient is multiplied by the original particle velocity and then the chaotic perturbation is added to update the velocity. The particle position is updated based on the updated velocity, and the fitness function value of the new position is calculated. The position of the globally optimal particle is recorded.

9. The remote data acquisition and control system for a spray product filling production line according to claim 1, characterized in that, The step of sending the remote control commands to the field actuator via industrial Ethernet specifically includes: A publish-subscribe model is adopted, and a message publishing node is established at the central control terminal. The message publishing node encapsulates remote control commands into data packets according to the OPC unified architecture protocol. The data packets are sent to the field industrial gateway via industrial Ethernet, and the field industrial gateway operates a subscription node to receive the data packets; The on-site industrial gateway parses the data packet to obtain the proportional valve opening value, heater power value, and conveyor motor speed value from the remote control command. The proportional valve opening value is converted into a voltage signal and output to the proportional valve through the analog output module. The heater power value is converted into a pulse width modulation signal and output to the heater through the relay module. The conveyor motor speed value is converted into a frequency signal and output to the frequency converter through the digital output module.

10. The remote data acquisition and control system for a spray product filling production line according to claim 9, characterized in that, The field industrial gateway has a built-in network interruption resume buffer. When the network connection with the central control terminal is interrupted, the received remote control commands are cached locally and executed in the order of timestamps.

Images