Environment-friendly high-glow high-impact pbt particles and preparation method thereof

Abstract

The present application relates to the field of PBT particle preparation, and particularly relates to an environment-friendly high-glow-wire high-impact PBT particle and a preparation method thereof, which comprises the following steps: feeding the actual feeding speed of the flame-retardant powder, the actual feeding speed of the glass fiber and the discharging speed of the equipment into a multi-scale CNN and an attention mechanism architecture to generate a feeding feature vector; feeding the feeding feature vector into an LSTM model to capture the feeding fluctuation law and generate a feeding bridging and caking feature; feeding the feeding bridging and caking feature into a full connection network to generate a predicted powder feeding speed and a predicted fiber feeding speed, and adjusting the expected powder feeding rotating speed and the expected fiber feeding rotating speed of the forced feeding equipment. Through the multi-model collaborative architecture, the feeding speed capable of actively predicting the feeding fluctuation and inhibiting the bridging and caking is generated, so as to ensure the uniformity of the component dispersion of the feeding, and realize the stable preparation performance of the environment-friendly high-glow-wire high-impact PBT particle.

Description

Technical Field

[0001] This invention relates to the field of PBT particle preparation, and in particular to an environmentally friendly high-heat-wire, high-impact PBT particle and its preparation method. Background Technology

[0002] Polybutylene terephthalate (PBT), as a high-performance engineering plastic, possesses excellent mechanical properties, heat resistance, and moldability, and is widely used in electronics, automotive manufacturing, and other fields. With downstream industries increasingly demanding higher safety and reliability from materials, stringent requirements are being placed on the high glow wire flame retardancy and impact resistance of PBT granules.

[0003] For environmentally friendly high-glow wire high-impact PBT granules that do not contain the harmful component triphenyl phosphate (TPP), most of them contain composite flame retardants and glass fibers. The composite flame retardant system achieves high-efficiency flame retardancy, while the glass fibers can improve the material's rigidity and dimensional stability, and improve its impact performance.

[0004] However, in the process of adding the composite flame retardant and glass fiber to PBT resin to produce PBT granules, the accuracy of the feeding speed setting of the forced feeding equipment determines the uniformity of component dispersion, thus becoming a key bottleneck affecting the final performance of PBT granules. Specifically, the composite flame retardant is mostly in powder form, while the glass fiber is in chopped filaments. During forced feeding, both are prone to bridging (clogging of the feeding channel) or agglomeration (powder agglomeration) due to the material characteristics of powder agglomeration and fiber entanglement, as well as fluctuations in equipment parameters. This leads to reduced uniformity of component dispersion in the feeding mixture, and consequently, reduced flame retardant and impact resistance of the PBT granules.

[0005] Currently, the feeding control of forced feeding equipment largely relies on real-time feedback adjustment from loss-in-weight weighing, adjusting the feeding speed by comparing the actual feeding rate with the target rate. However, this process is a passive response, unable to predict short-term fluctuations such as sudden powder agglomeration, and difficult to capture the potential impact of changes in material properties on feeding. This results in a long lag time for feeding speed adjustment, especially in high-capacity continuous production with forced feeding equipment. This can easily lead to large batch-to-batch performance variations in PBT granules, thus affecting the product qualification rate for downstream high-performance applications.

[0006] Therefore, how to actively predict feeding fluctuations and adjust the feeding speed of composite flame retardants and glass fibers to suppress cross-linking and agglomeration during the preparation of environmentally friendly high-glow-wire, high-impact PBT granules, thereby ensuring the uniformity of component dispersion and achieving stable product performance of PBT granules, is a technical problem that needs to be solved. Summary of the Invention

[0007] To this end, the present invention provides an environmentally friendly high-glow-wire, high-impact PBT particle and its preparation method. Through a multi-model collaborative architecture of CNN, attention mechanism, and LSTM, the invention realizes the short-term key characteristics based on the actual feeding speed of composite flame retardant and glass fiber during the PBT particle preparation process, identifies the characteristic signals of cross-bridging and agglomeration, and then maps and generates a feeding speed that can actively predict feeding fluctuations and suppress cross-bridging and agglomeration, thereby ensuring the uniformity of component dispersion and achieving stable preparation performance of environmentally friendly high-glow-wire, high-impact PBT particles.

[0008] To achieve the above objectives, this invention proposes a method for preparing environmentally friendly high-heat-wire, high-impact PBT particles, comprising:

[0009] Composite flame retardant powder and glass fiber are added to a forced feeding device to force-feed PBT resin, and then PBT granule premix is ​​discharged. The PBT granule premix is ​​extruded and injection molded to obtain PBT granules.

[0010] The forced feeding device obtains the actual feeding speed of flame retardant powder, the actual feeding speed of glass fiber, and the equipment discharge speed through multiple loss-in-weight scales.

[0011] The actual feeding speed of the flame retardant powder, the actual feeding speed of the glass fiber, and the discharge speed of the equipment are used to perform short-term feature extraction using a feature extraction model based on a multi-scale CNN and attention mechanism architecture to generate a feeding feature vector.

[0012] The feeding feature vector is used to capture the feeding pattern through an LSTM model to generate feeding bridging and agglomeration features;

[0013] The feed bridging and agglomeration characteristics are passed through a fully connected network to generate predicted powder feed rate and predicted fiber feed rate, and the desired powder feed speed and desired fiber feed speed of the forced feeder are adjusted based on the predicted powder feed rate and predicted fiber feed rate.

[0014] Furthermore, the feature extraction model includes a powder feature extraction branch and a fiber feature extraction branch, and the feeding feature vector includes a powder feeding feature vector and a fiber feeding feature vector. The process of generating the feeding feature vector through short-term feature extraction using the feature extraction model includes:

[0015] The actual feeding speed of the flame retardant powder, the target feeding speed of the powder, and the actual operating data of the feeder are used as the powder branch input data. The powder branch input data is then passed through the powder multi-scale convolutional layer of the powder feature extraction branch to generate multi-scale powder feeding features.

[0016] The multi-scale powder feeding features and the powder branch input data are weighted by multi-scale feature guidance through the powder attention mechanism of the powder feature extraction branch to generate the powder feeding feature vector.

[0017] The actual feeding speed of the glass fiber, the target feeding speed of the fiber, and the actual operating data of the feeder are used as fiber branch input data. The fiber branch input data are then processed through the fiber feature extraction branch multi-scale convolutional layer to generate multi-scale fiber feeding features.

[0018] The multi-scale fiber feeding features and fiber branch input data are weighted by multi-scale feature guidance through the fiber attention mechanism of the fiber feature extraction branch to generate the fiber feeding feature vector.

[0019] Furthermore, the multi-scale powder feeding characteristics include powder fluctuation characteristics, powder bridging trend characteristics, and powder flow characteristics. The process of generating the powder feeding feature vector through multi-scale convolutional layers and a powder attention mechanism includes:

[0020] The powder branch input data is passed through the first powder convolutional layer, the second powder convolutional layer and the third powder convolutional layer of the powder multi-scale convolutional layer respectively to generate powder fluctuation characteristics, powder bridging trend characteristics and powder flow characteristics in sequence;

[0021] The standard deviation, skewness, and kurtosis of the powder branch input data are respectively processed through the powder attention mechanism to generate fluctuation attention weight, agglomeration attention weight, and bridging attention weight. The fluctuation attention weight, agglomeration attention weight, and bridging attention weight are then weighted and summed with the powder fluctuation characteristics, powder bridging trend characteristics, and powder flow characteristics to generate the powder feeding feature vector.

[0022] The multi-scale fiber feeding features include fiber breakage features, fiber entanglement trend features, and fiber transport features. The process of generating fiber feeding feature vectors through fiber multi-scale convolutional layers and fiber attention mechanisms includes:

[0023] The fiber branch input data is passed through the first fiber convolutional layer, the second fiber convolutional layer, and the third fiber convolutional layer of the fiber multi-scale convolutional layer respectively to generate fiber breakage characteristics, fiber entanglement trend characteristics, and fiber transport characteristics in sequence.

[0024] The mean, variance, and information entropy of the fiber branch input data are respectively processed through the fiber attention mechanism to generate breakage attention weight, entanglement attention weight, and feeding attention weight. The breakage attention weight, entanglement attention weight, and feeding attention weight are then weighted and summed with the fiber breakage feature, fiber entanglement trend feature, and fiber transport feature to generate the fiber feeding feature vector.

[0025] The convolution kernel sizes of the convolutional layers, from smallest to largest, are as follows: the first powder convolutional layer, the first fiber convolutional layer, the second powder convolutional layer, the second fiber convolutional layer, the third powder convolutional layer, and the third fiber convolutional layer.

[0026] Furthermore, the LSTM model includes a fast LSTM branch for powder, a slow LSTM branch for fibers, and an interaction layer. The process of generating feed bridging and agglomeration features through the LSTM model includes:

[0027] The powder feeding feature vector and its change are respectively passed through the powder fast LSTM branch with fiber interaction input gate and change rate enhanced forget gate to generate the powder hidden state;

[0028] The fiber feeding feature vector and its changes are passed through a slow LSTM branch with a trend-enhancing input gate and a smooth forgetting gate to generate the fiber hidden state.

[0029] The powder hidden state and fiber hidden state are passed through an interaction layer to generate the feed bridging agglomeration feature.

[0030] Furthermore, the process of generating the hidden state of the powder includes:

[0031] The powder hidden state, the powder feeding feature vector, and the spliced ​​vector of the fiber hidden state from the previous time step are linearly mapped to generate the fiber interaction input gate and the initial forget gate. The differences of the powder feeding feature vectors from multiple time steps are weighted and summed on the initial forget gate to generate the rate of change enhanced forget gate. The powder hidden state at the current time step is updated based on the fiber interaction input gate and the rate of change enhanced forget gate.

[0032] The process of generating the hidden state of fibers includes:

[0033] The glass fiber feeding trend value is calculated based on the difference between the actual feeding speed and the target feeding speed of the glass fiber within a set time window. The initial input gate based on the fiber feeding feature vector and the glass fiber feeding trend value are weighted and summed to generate the trend enhancement input gate. The initial forgetting gate based on the fiber feeding feature vector is weighted and summed with the rate of change of the fiber feeding feature vector to generate the smooth forgetting gate. The hidden state of the fiber at the current time step is updated based on the trend enhancement input gate and the smooth forgetting gate.

[0034] Furthermore, the process of generating feed bridging cluster features through the interaction layer includes:

[0035] The powder hidden state and the fiber hidden state are weighted and summed to generate a linear term;

[0036] The interaction term is obtained by multiplying the powder hidden state and the fiber hidden state element by element.

[0037] The linear term and the interaction term are weighted and summed to generate a comprehensive hidden state, and the comprehensive hidden state, the powder hidden state, and the fiber hidden state at the final time step are used as the feed bridging and agglomeration features.

[0038] Furthermore, the process of adjusting the desired powder feed speed and the desired fiber feed speed includes:

[0039] Based on the difference between the predicted powder feeding speed and the target powder feeding speed, a powder adjustment amount is generated. Based on the powder adjustment amount and the calculation of the current powder feeding speed, a desired powder feeding speed is generated.

[0040] The fiber adjustment amount is generated based on the difference between the predicted fiber feeding speed and the target fiber feeding speed, and the desired fiber feeding speed is generated based on the calculation of the fiber adjustment amount and the current fiber feeding speed.

[0041] Furthermore, the preparation method further includes:

[0042] Based on the mean square error between the predicted powder feeding rate and the optimal powder feeding rate of the sample, and the mean square error between the predicted fiber feeding rate and the optimal fiber feeding rate of the sample, a main loss term is constructed.

[0043] A flow stability loss term is constructed based on the ratio of the predicted powder feeding rate to the predicted fiber feeding rate and the ratio of the theoretical powder flow rate to the theoretical fiber flow rate.

[0044] A comprehensive loss function is constructed based on the weighted sum of the main loss term and the traffic stability loss term. The feature extraction model, the LSTM model, and the fully connected network are then collaboratively optimized and trained using the comprehensive loss function.

[0045] In particular, the intelligent feeding control based on multi-scale CNN, attention mechanism, LSTM, and fully connected network solves the problem of imbalance caused by powder agglomeration and fiber entanglement in traditional feeding. The dual-branch multi-scale CNN combined with the attention mechanism accurately captures the powder agglomeration trend, flow stability, fiber entanglement risk, and breakage characteristics. Through attention weighting, the feature recognition accuracy is improved. The dual-branch LSTM is used to handle instantaneous changes in powder and long-term trends in fibers. The deviation between the prediction speed and the target value output by the fully connected network is controlled to achieve proactive correction, ensuring the uniformity of the feed components and significantly improving the batch-to-batch consistency of product quality.

[0046] The present invention also provides an environmentally friendly high glow wire high impact PBT granules prepared by the above method, comprising, by weight, 30-50 parts PBT resin, 5-40 parts glass fiber and 23.9-60.4 parts composite flame retardant, wherein the composite flame retardant comprises halogen flame retardant, nitrogen-based MCA flame retardant and auxiliary additives.

[0047] The halogenated flame retardant, by weight, comprises 8-15 parts brominated epoxy, 5-15 parts brominated polystyrene and 1-5 parts antimony trioxide;

[0048] The nitrogen-based MCA flame retardant comprises, by weight, 3-10 parts nitrogen-based flame retardant MCA, 2-6 parts stannate and 1-3 parts silicate.

[0049] Further, the auxiliary additives, by weight, include 3-5 parts toughening agent, 0.2-0.4 parts antioxidant, 0.5 parts dispersant and 0.2-0.5 parts coupling agent.

[0050] In particular, PBT granules achieve high glow wire and low smoke toxicity through halogenated flame retardants, nitrogen-based MCA flame retardants and auxiliary additives, ensuring that the formulation design takes into account both the performance of high glow wire and high impact and environmental protection characteristics.

[0051] Compared with the prior art, the beneficial effects of the present invention are that, through a multi-model collaborative architecture of CNN, attention mechanism and LSTM, the present invention realizes the short-term key features of the actual feeding speed of composite flame retardant and glass fiber in the PBT particle preparation process, identifies the characteristic signals of cross-bridging and agglomeration, and then maps and generates a feeding speed that can actively predict feeding fluctuations and suppress cross-bridging and agglomeration, thereby ensuring the uniformity of component dispersion and achieving stable preparation performance of environmentally friendly high glow wire and high impact PBT particles.

[0052] In particular, this invention, based on intelligent feeding control using multi-scale CNN, attention mechanism, LSTM, and fully connected network, solves the problem of imbalance caused by powder agglomeration and fiber entanglement in traditional feeding. The dual-branch multi-scale CNN combined with the attention mechanism accurately captures powder agglomeration trends, flow stability, and fiber entanglement risks and breakage characteristics. Through attention weighting, it improves the accuracy of feature recognition. The dual-branch LSTM specifically handles instantaneous changes in powder and long-term trends in fibers. The prediction speed and target value deviation control of the fully connected network output achieve proactive correction, ensuring the uniformity of component dispersion in the feed and significantly improving batch-to-batch consistency of product quality.

[0053] In particular, the PBT particles of the present invention achieve high glow wire and low smoke toxicity through halogenated flame retardants, nitrogen-based MCA flame retardants and auxiliary additives, ensuring that the formulation design takes into account both the performance of high glow wire and high impact and environmental protection characteristics. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the process for preparing environmentally friendly high-heat-wire high-impact PBT particles and its preparation method according to an embodiment of the present invention.

[0055] Figure 2 This is a schematic diagram of the powder feature extraction branch process of the environmentally friendly high-heat-wire high-impact PBT particles and their preparation method according to an embodiment of the present invention.

[0056] Figure 3 This is a schematic diagram of the LSTM model flow of the environmentally friendly high-glow wire high-impact PBT particles and their preparation method according to an embodiment of the present invention.

[0057] Figure 4 This is a schematic diagram of the overall process for preparing environmentally friendly high-glow wire high-impact PBT particles according to an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0059] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0060] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0061] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0062] like Figures 1 to 4 As shown, this invention provides an environmentally friendly high-glow-wire, high-impact PBT particle and its preparation method. Through a multi-model collaborative architecture of CNN, attention mechanism, and LSTM, it realizes the short-term key features based on the actual feeding speed of composite flame retardant and glass fiber during the PBT particle preparation process, identifies the characteristic signals of cross-bridging and agglomeration, and then maps and generates a feeding speed that can actively predict feeding fluctuations and suppress cross-bridging and agglomeration, thereby ensuring the uniformity of component dispersion and achieving stable preparation performance of environmentally friendly high-glow-wire, high-impact PBT particles.

[0063] like Figure 1 As shown, this embodiment proposes a method for preparing environmentally friendly high-glow wire, high-impact PBT particles. The PBT particles are composed of PBT resin, glass fiber, and a composite flame retardant. The preparation method of the PBT particles includes:

[0064] Composite flame retardant powder and glass fiber are added to a forced feeding device to force-feed PBT resin, and then PBT granule premix is ​​discharged. The PBT granule premix is ​​extruded and injection molded to obtain PBT granules.

[0065] The forced feeding device obtains the actual feeding speed of flame retardant powder, the actual feeding speed of glass fiber, and the equipment discharge speed through multiple loss-in-weight scales.

[0066] The actual feeding speed of the flame retardant powder, the actual feeding speed of the glass fiber, and the discharge speed of the equipment are used to perform short-term feature extraction using a feature extraction model based on a multi-scale CNN and attention mechanism architecture to generate a feeding feature vector.

[0067] The feeding feature vector is used to capture the feeding pattern through an LSTM model to generate feeding bridging and agglomeration features;

[0068] The feed bridging and agglomeration characteristics are passed through a fully connected network to generate predicted powder feed rate and predicted fiber feed rate, and the desired powder feed speed and desired fiber feed speed of the forced feeder are adjusted based on the predicted powder feed rate and predicted fiber feed rate.

[0069] In particular, feature extraction models utilizing multi-scale CNN (Convolutional Neural Network) and attention mechanisms can extract short-term, material-compatible feeding features for the composite flame retardant powder's tendency to agglomerate and its highly variable flowability, and for the glass fiber's tendency to entangle and its uniform density. LSTM can extract long-term, material-compatible feeding bridging and agglomeration features for the composite flame retardant powder's unstable flowability and for the glass fiber's need for continuous and smooth feeding. Furthermore, the extracted feeding bridging and agglomeration features can be mapped through an MLP architecture output mapping model to generate predicted powder feeding speeds that better reflect the composite flame retardant powder and emphasize short-term fluctuations, and predicted fiber feeding speeds that better reflect the glass fiber and emphasize long-term stability. This enables more precise predictive control of the forced feeding equipment's feeding speed, reducing the probability of bridging and agglomeration, ensuring uniform component dispersion after mixing glass fiber and composite flame retardant with the main PBT tree material in the forced feeding equipment, and achieving stable production performance of environmentally friendly, high-glow-wire, high-impact PBT granules.

[0070] Furthermore, the feature extraction model includes a powder feature extraction branch and a fiber feature extraction branch, and the feeding feature vector includes a powder feeding feature vector and a fiber feeding feature vector. The process of generating the feeding feature vector through short-term feature extraction using the feature extraction model includes:

[0071] The actual feeding speed of the flame retardant powder, the target feeding speed of the powder, and the actual operating data of the feeder are used as the powder branch input data. The powder branch input data is then passed through the powder multi-scale convolutional layer of the powder feature extraction branch to generate multi-scale powder feeding features.

[0072] The multi-scale powder feeding features and the powder branch input data are weighted by multi-scale feature guidance through the powder attention mechanism of the powder feature extraction branch to generate the powder feeding feature vector.

[0073] The actual feeding speed of the glass fiber, the target feeding speed of the fiber, and the actual operating data of the feeder are used as fiber branch input data. The fiber branch input data are then processed through the fiber feature extraction branch multi-scale convolutional layer to generate multi-scale fiber feeding features.

[0074] The multi-scale fiber feeding features and fiber branch input data are weighted by multi-scale feature guidance through the fiber attention mechanism of the fiber feature extraction branch to generate the fiber feeding feature vector.

[0075] Preferably, the actual operating data of the feeder includes the motor speed of the feeding mixture and the equipment discharge speed.

[0076] In particular, by focusing on the agglomeration and bridging characteristics of powder through powder branching, and adapting to the entanglement and breakage characteristics of fibers through fiber branching, the characteristics of powder and fiber are avoided from being confused with each other.

[0077] Furthermore, the multi-scale convolutional layer for powder includes a first powder convolutional layer, a second powder convolutional layer, and a third powder convolutional layer. The multi-scale powder feeding features include powder fluctuation features, powder bridging trend features, and powder flow features. The process of generating the powder feeding feature vector through the multi-scale convolutional layer and the powder attention mechanism includes:

[0078] The powder branch input data is passed through the first powder convolutional layer, the second powder convolutional layer, and the third powder convolutional layer, respectively, to generate powder fluctuation characteristics, powder bridging trend characteristics, and powder flow characteristics;

[0079] The standard deviation, skewness, and kurtosis of the powder branch input data are respectively processed through the powder attention mechanism to generate fluctuation attention weight, agglomeration attention weight, and bridging attention weight. The fluctuation attention weight, agglomeration attention weight, and bridging attention weight are then weighted and summed with the powder fluctuation characteristics, powder bridging trend characteristics, and powder flow characteristics to generate the powder feeding feature vector in sequence.

[0080] The fiber multi-scale convolutional layer includes a first fiber convolutional layer, a second fiber convolutional layer, and a third fiber convolutional layer. The multi-scale fiber feeding features include fiber breakage features, fiber entanglement trend features, and fiber transport features. The process of generating fiber feeding feature vectors through the fiber multi-scale convolutional layer and fiber attention mechanism includes:

[0081] The fiber branch input data is passed through the first fiber convolutional layer, the second fiber convolutional layer, and the third fiber convolutional layer respectively to generate fiber breakage characteristics, fiber entanglement trend characteristics, and fiber transport characteristics in sequence.

[0082] The mean, variance, and information entropy of the fiber branch input data are respectively processed through the fiber attention mechanism to generate breakage attention weight, entanglement attention weight, and feeding attention weight. The breakage attention weight, entanglement attention weight, and feeding attention weight are then weighted and summed with the fiber breakage feature, fiber entanglement trend feature, and fiber transport feature to generate the fiber feeding feature vector.

[0083] The convolution kernel sizes of the convolutional layers, from smallest to largest, are as follows: the first powder convolutional layer, the first fiber convolutional layer, the second powder convolutional layer, the second fiber convolutional layer, the third powder convolutional layer, and the third fiber convolutional layer.

[0084] Specifically, the powder multi-scale convolutional layer can be represented as:

[0085]

[0086] In the formula, These represent powder fluctuation characteristics, powder bridging trend characteristics, and powder flow characteristics, respectively. express Activation function The three convolutional kernels represent the first (3x1), second (7x1), and third (15x1) convolutional kernels of the powder feature extraction branch p, respectively, covering feeding data within 0.6 seconds, 1.4 seconds, and 3 seconds. This represents the input data for the powder branch of the powder feature extraction branch p. The input data for the powder branch includes the actual feeding speed of the flame retardant powder, the target feeding speed of the powder, and the actual operating data of the feeder. These represent the 3x1 convolution kernel bias term of the first powder convolutional layer, the 7x1 convolution kernel bias term of the second powder convolutional layer, and the 15x1 convolution kernel bias term of the third powder convolutional layer, respectively, in the powder feature extraction branch p.

[0087] In particular, multi-scale features were extracted to address the tendency of composite flame retardant powders to agglomerate through multi-scale convolutional layers.

[0088] Specifically, the powder attention mechanism can be represented as:

[0089]

[0090] In the formula, This represents the attention weight vector of the powder attention mechanism p, containing 3 elements. express Normalization function, Let represent the learnable weight matrix and learnable bias term of the powder attention mechanism p, respectively. These represent the standard deviation, skewness, and kurtosis of the powder branch input data, respectively. The powder branch input data includes the actual feeding rate of the flame retardant powder, the target feeding rate of the powder, and the actual operating data of the feeder. The standard deviation reflects the degree of fluctuation in the powder feeding data by statistically analyzing the dispersion of the actual feeding rate of the flame retardant powder. The skewness reflects the degree of typical agglomeration precursors by statistically analyzing the short-term sharp drop and slow recovery of the powder feeding rate. The kurtosis reflects the instantaneous impact of agglomeration on the feeding mechanism. This represents the feature vector of powder feeding. The attention weight vector consists of three elements: fluctuation attention weight, clustering attention weight, and bridging attention weight. These represent the characteristics of powder fluctuation, powder bridging trend, and powder flow, respectively.

[0091] In particular, the powder attention mechanism assigns different importance weights to multi-scale features based on the different characteristics reflected in the powder input data, enabling the feature extraction model to learn to focus on more important information.

[0092] Specifically, the powder multi-scale convolutional layer can be represented as:

[0093]

[0094] In the formula, These represent the fiber breakage characteristics, fiber entanglement trend characteristics, and fiber transport characteristics of fiber feature extraction branch f, respectively. express Activation function The first fiber convolutional layer kernel (5x1 size), the second fiber convolutional layer kernel (11x1 size), and the third fiber convolutional layer kernel (21x1 size) of the fiber feature extraction branch f represent the feeding data within 0.6 seconds, 1.4 seconds, and 3 seconds, respectively. This represents the fiber branch input data for the fiber feature extraction branch f. The fiber branch input data includes the actual glass fiber feeding speed, the target fiber feeding speed, and the actual operating data of the feeder. These represent the 5x1 convolution kernel bias term of the first fiber convolutional layer, the 11x1 convolution kernel bias term of the second powder fiber layer, and the 21x1 convolution kernel bias term of the third fiber convolutional layer, respectively, for the fiber feature extraction branch f.

[0095] Specifically, the fiber attention mechanism can be represented as:

[0096]

[0097] In the formula, This represents the attention weight vector of the fiber attention mechanism f, containing 3 elements. express Normalization function, Let f represent the learnable weight matrix and learnable bias term of the fiber attention mechanism f, respectively. Let $\mathbf$, $\mathbf$, $\mathbf$, and $\mathbf$ represent the mean, variance, and information entropy of the fiber branch input data. The fiber branch input data includes the actual fiber feeding rate, the target fiber feeding rate, and the actual operating data of the feeder. The mean measures the stability of the average fiber feeding rate to reflect whether there is feeding obstruction caused by fiber entanglement. The variance measures the fluctuation of the fiber feeding rate to reflect whether there is local entanglement or breakage of the fiber. The information entropy measures the disorder of the fiber feeding data and is calculated using the classical formula for information entropy based on discrete probability. This classical formula is: ,in This represents the probability that the data falls within interval k. This represents the fiber feed feature vector. The attention weight vector consists of three elements: the break attention weight, the entanglement attention weight, and the feeding attention weight. These represent the fiber breakage characteristics, fiber entanglement trend characteristics, and fiber transport characteristics of the fiber feature extraction branch f, respectively.

[0098] In particular, the powder attention mechanism enables the assignment of different importance weights to multi-scale features based on the different characteristics reflected in the fiber input data, allowing the feature extraction model to learn to focus on more important information.

[0099] Furthermore, the LSTM model includes a fast LSTM branch for powder, a slow LSTM branch for fibers, and an interaction layer. The process of generating feed bridging and agglomeration features through the LSTM model includes:

[0100] The powder feeding feature vector and its change are respectively passed through the powder fast LSTM branch with fiber interaction input gate and change rate enhanced forget gate to generate the powder hidden state;

[0101] The fiber feeding feature vector and its changes are passed through a slow LSTM branch with a trend-enhancing input gate and a smooth forgetting gate to generate the fiber hidden state.

[0102] The powder hidden state and fiber hidden state are passed through an interaction layer to generate the feed bridging agglomeration feature.

[0103] In particular, by addressing the temporal dynamic differences and material coupling characteristics of anomalies in powder and glass fiber feeding, branch function differentiation and interactive fusion are used to achieve comprehensive capture of the temporal characteristics of complex feeding systems. This enables precise modeling of the instantaneous evolution of powder agglomeration, the development of fiber entanglement, and material linkage anomalies, providing highly accurate feature support for early warning and dynamic control of feeding bridging agglomeration.

[0104] Furthermore, the process of generating the hidden state of the powder includes:

[0105] The powder hidden state, the powder feeding feature vector, and the spliced ​​vector of the fiber hidden state from the previous time step are linearly mapped to generate the fiber interaction input gate and the initial forget gate. The differences of the powder feeding feature vectors from multiple time steps are weighted and summed on the initial forget gate to generate the rate of change enhanced forget gate. The powder hidden state at the current time step is updated based on the fiber interaction input gate and the rate of change enhanced forget gate.

[0106] The process of generating the hidden state of fibers includes:

[0107] The glass fiber feeding trend value is calculated based on the difference between the actual feeding speed and the target feeding speed of the glass fiber within a set time window. The initial input gate based on the fiber feeding feature vector and the glass fiber feeding trend value are weighted and summed to generate the trend enhancement input gate. The initial forgetting gate based on the fiber feeding feature vector is weighted and summed with the rate of change of the fiber feeding feature vector to generate the smooth forgetting gate. The hidden state of the fiber at the current time step is updated based on the trend enhancement input gate and the smooth forgetting gate.

[0108] In particular, the refined LSTM architecture, which features dynamic adaptation through gate mechanism, cross-material information fusion, and enhanced temporal features, enables the two branches to more accurately capture the temporal feature patterns of their respective materials, while strengthening the perception of cross-material correlations.

[0109] Specifically, the fast LSTM branching for powder can be represented as:

[0110]

[0111] In the formula, These represent the fiber interaction input gate, rate-of-change enhanced forgetting gate, output gate, candidate cell state, current cell state update, previous time step cell state update, and current time step hidden powder state, respectively, of the fast LSTM branch for powder. This represents the Sigmoid function. express function, This indicates element-wise multiplication. These represent the weight matrices for the input gate, forget gate, output gate, and candidate cell states of the fast LSTM branch for powder, respectively. These represent the input gate, forget gate, output gate, and bias term for the candidate cell state in the fast LSTM branch for powder, respectively. Indicates the hidden state of the powder in the previous time step. The powder feeding feature vector at time step t The fiber hiding state of the previous time step The concatenated vector, Represents the weighting coefficient for the rate of change, where Indicates the initial forget gate, The rate of change of the characteristic vector of powder feeding at multiple time steps is preferably obtained by averaging the pairwise differences of the three closest time steps.

[0112] Specifically, the fiber fast LSTM branching can be represented as:

[0113]

[0114] In the formula, These represent the trend-enhancing input gate, smoothing forgetting gate, output gate, candidate cell state, current cell state update, previous time step cell state update, previous time step cell state update, current time step fiber hidden state, and previous time step fiber hidden state, respectively, for the fiber fast LSTM branching. This represents the Sigmoid function. express function, This indicates element-wise multiplication. Let represent the weight matrices of the input gate, forget gate, output gate, and candidate cell states of the fast LSTM branch, respectively. represents the input gate, forget gate, output gate, and bias term of the candidate cell state in the fast LSTM branch, respectively, and [ ] represents the feature vector concatenation operation. These represent the trend-enhancing weighting coefficient, the change-reducing weighting coefficient, and the historical cell state preservation coefficient, respectively. This indicates the glass fiber feeding trend value. Represents the fiber feed feature vector The rate of change of , where Indicates the initial input gate. This indicates the initial forgetting gate of the fiber.

[0115] The calculation process for the glass fiber feeding trend value can be expressed as follows:

[0116]

[0117] In the formula, This indicates the glass fiber feeding trend value. This indicates the set time window, preferably 5. This indicates the actual feeding speed of the glass fiber. This indicates the target fiber feeding rate.

[0118] Furthermore, the process of generating feed bridging cluster features through the interaction layer includes:

[0119] The powder hidden state and the fiber hidden state are weighted and summed to generate a linear term;

[0120] The interaction term is obtained by multiplying the powder hidden state and the fiber hidden state element by element.

[0121] The linear term and the interaction term are weighted and summed to generate a comprehensive hidden state, and the comprehensive hidden state, the powder hidden state, and the fiber hidden state at the final time step are used as the feed bridging and agglomeration features.

[0122] In particular, by deeply fusing the hidden states of the fast LSTM branch output for powder and the slow LSTM branch output for fiber, the resulting feed bridging and agglomeration features combine information on instantaneous anomalies of powder, long-term trend patterns of fiber, and cross-material linkage characteristics, providing highly recognizable feature support for dynamic feed control.

[0123] Specifically, the interaction layer can be represented as:

[0124]

[0125] In the formula, This represents the overall hidden state at the current time step. This indicates the set time window, preferably 5. This indicates the actual feeding speed of the glass fiber. This indicates the target fiber feeding rate.

[0126] Specifically, the feed bridging and agglomeration features are processed through a fully connected network to generate predicted powder feed rate and predicted fiber feed rate. The fully connected network includes a ReLU activation function convolutional layer, a Dropout layer, a residual connection layer, and a LayerNorm layer.

[0127] Furthermore, the process of adjusting the desired powder feed speed and the desired fiber feed speed includes:

[0128] Based on the difference between the predicted powder feeding speed and the target powder feeding speed, a powder adjustment amount is generated. Based on the powder adjustment amount and the calculation of the current powder feeding speed, a desired powder feeding speed is generated.

[0129] The fiber adjustment amount is generated based on the difference between the predicted fiber feeding speed and the target fiber feeding speed, and the desired fiber feeding speed is generated based on the calculation of the fiber adjustment amount and the current fiber feeding speed.

[0130] In particular, the closed-loop control mechanism of the feeding speed avoids the limitations of traditional fixed speed or experience-based adjustments, and realizes active, dynamic and precise control of the feeding speed of the feeder.

[0131] Specifically, the process of generating the desired powder feed speed and the desired fiber feed speed can be expressed as:

[0132]

[0133] In the formula, These represent the expected powder feed speed and the expected fiber feed speed at the next moment, respectively. These represent the conversion coefficients for the two feeding speeds and the motor speed, respectively. These represent the conversion coefficients for the two feeding speeds and the motor speed, respectively. These represent the predicted powder feed rate and the predicted fiber feed rate, respectively. These represent the target feed rate for powder and the target feed rate for fiber, respectively. These represent the current powder feeding speed and the current fiber feeding speed, respectively.

[0134] Furthermore, the preparation method further includes:

[0135] Based on the mean square error between the predicted powder feeding rate and the optimal powder feeding rate of the sample, and the mean square error between the predicted fiber feeding rate and the optimal fiber feeding rate of the sample, a main loss term is constructed.

[0136] A flow stability loss term is constructed based on the ratio of the predicted powder feeding rate to the predicted fiber feeding rate and the ratio of the theoretical powder flow rate to the theoretical fiber flow rate.

[0137] A comprehensive loss function is constructed based on the weighted sum of the main loss term and the traffic stability loss term. The feature extraction model, the LSTM model, and the fully connected network are then collaboratively optimized and trained using the comprehensive loss function.

[0138] In particular, the flow stability loss term is constructed by predicting the deviation between the ratio of powder to fiber velocity and the theoretical flow ratio of powder to fiber (formulation requirements). This forces the model to maintain the material ratio within the range required by the formulation and keep it stable during prediction (without agglomeration or bridging), thus avoiding product quality problems caused by imbalance in the ratio and ensuring the quality of products fed by forced feeding.

[0139] Specifically, the process of generating the desired powder feed speed and the desired fiber feed speed can be expressed as:

[0140]

[0141] In the formula, This represents the comprehensive loss function, where N represents the total number of samples. These represent the weighting coefficients of the main loss term and the flow stability loss term, respectively. These represent the predicted powder feeding rate and the predicted fiber feeding rate for the i-th input sample, respectively. This represents the optimal feeding rate for the actual powder and the optimal feeding rate for the actual fiber of the i-th input sample. These represent the bulk densities of the composite flame retardant and glass fiber, respectively, which are determined based on the formulation ratio of PTB particles. These represent the cross-sectional areas of the composite flame retardant and the glass fiber feed port of the forced feeding equipment, respectively.

[0142] This embodiment also provides an environmentally friendly high glow wire high impact PBT particle prepared by the above method. The PBT particle is composed of 30-50 parts of PBT resin, 5-40 parts of glass fiber and 23.9-60.4 parts of composite flame retardant by weight. The composite flame retardant includes halogen flame retardant, nitrogen-based MCA flame retardant and auxiliary additives.

[0143] The halogenated flame retardant, by weight, comprises 8-15 parts brominated epoxy, 5-15 parts brominated polystyrene and 1-5 parts antimony trioxide;

[0144] The nitrogen-based MCA flame retardant comprises, by weight, 3-10 parts nitrogen-based flame retardant MCA, 2-6 parts stannate and 1-3 parts silicate.

[0145] In particular, in the original high-glow wire PBT formulation, the bromine-antimony flame retardant system required the combination of triphenyl phosphate (TPP) and nitrogen-based flame retardants to form carbon and isolate oxygen in the gas phase, preventing PBT from producing open flames at 750 degrees Celsius. However, TPP is currently considered environmentally unfriendly and harmful to human health. Therefore, the PBT particles in this embodiment achieve environmentally friendly high-glow wire with high impact performance by eliminating TPP.

[0146] Further, the auxiliary additives, by weight, include 3-5 parts toughening agent, 0.2-0.4 parts antioxidant, 0.5 parts dispersant and 0.2-0.5 parts coupling agent.

[0147] Specifically, the process of preparing composite flame retardant powder from halogenated flame retardants, nitrogen-based MCA flame retardants, and auxiliary additives includes: first drying PBT resin, halogenated flame retardants, nitrogen-based MCA flame retardants, and auxiliary additives in an oven at 110 degrees Celsius for 3 hours, and then testing the moisture content with a moisture meter to ensure it remains below 0.1%; then stirring the coupling agent and flame retardant powder in a pulverizer to achieve pre-dispersion; and finally, thoroughly mixing all materials of the halogenated flame retardants, nitrogen-based MCA flame retardants, and auxiliary additives to form composite flame retardant powder.

[0148] Composite flame retardant powder, PBT resin, and glass fiber are fed into PBT granule premix using a forced feeding device. The process of preparing PBT granules from the PBT granule premix includes: extruding the PBT granule premix through a twin-screw extruder at a temperature of 225-255 degrees Celsius, stretching and rinsing with water (cold dewatering tank), drying (water blowing and descaling machine), pelletizing (granulator), sieving (vibrating screen), drying (oven), injection molding (injection molding machine and mold), and testing (glow wire tester and impact tester) to produce PBT granules.

[0149] Specifically, the test results of the PBT particles' components are shown in the table below, "Inspection Specification for Flame-Retardant Reinforced Finished Products":

[0150]

[0151] In particular, PBT granules achieve high glow wire and low smoke toxicity through halogenated flame retardants, nitrogen-based MCA flame retardants and auxiliary additives, ensuring that the formulation design takes into account both the performance of high glow wire and high impact and environmental protection characteristics.

[0152] In this embodiment, a multi-model collaborative architecture of CNN, attention mechanism, and LSTM is used to realize the short-term key features of the actual feeding speed of composite flame retardant and glass fiber during PBT particle preparation. This identifies the characteristic signals of bridging and agglomeration, and then maps them to generate a feeding speed that can actively predict feeding fluctuations and suppress bridging and agglomeration, thereby ensuring the uniformity of component dispersion and achieving stable performance in the preparation of environmentally friendly, high-glow-wire, high-impact PBT particles. Intelligent feeding control based on multi-scale CNN, attention mechanism, LSTM, and fully connected networks solves the problem of imbalance caused by powder agglomeration and fiber entanglement in traditional feeding methods. A dual-branch multi-scale CNN combined with an attention mechanism accurately captures powder agglomeration trends, flow stability, and fiber entanglement risks and breakage characteristics. Attention weighting improves feature recognition accuracy. A dual-branch LSTM specifically handles instantaneous powder changes and long-term fiber trends. The deviation between the predicted speed and target value output by the fully connected network is controlled, enabling proactive correction and ensuring the uniformity of component dispersion in the feeding process, significantly improving batch-to-batch consistency of product quality. PBT granules achieve high glow wire and low smoke toxicity through halogenated flame retardants, nitrogen-based MCA flame retardants, and auxiliary additives, ensuring that the formulation design takes into account both the high glow wire and high impact performance and environmental protection characteristics.

[0153] Those skilled in the art will recognize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0154] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0155] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for preparing environmentally friendly high-heat-wire, high-impact PBT particles, characterized in that, include: Composite flame retardant powder and glass fiber are added to a forced feeding device to force-feed PBT resin, and then PBT granule premix is ​​discharged. The PBT granule premix is ​​extruded and injection molded to obtain PBT granules. The forced feeding device obtains the actual feeding speed of flame retardant powder, the actual feeding speed of glass fiber, and the equipment discharge speed through multiple loss-in-weight scales. The actual feeding speed of the flame retardant powder, the actual feeding speed of the glass fiber, and the discharge speed of the equipment are used to perform short-term feature extraction using a feature extraction model based on a multi-scale CNN and attention mechanism architecture to generate a feeding feature vector. The feeding feature vector is used to capture the feeding pattern through an LSTM model to generate feeding bridging and agglomeration features; The feed bridging and agglomeration characteristics are passed through a fully connected network to generate predicted powder feed rate and predicted fiber feed rate, and the desired powder feed speed and desired fiber feed speed of the forced feeder are adjusted based on the predicted powder feed rate and predicted fiber feed rate.

2. The method for preparing environmentally friendly high-heat-wire, high-impact PBT particles according to claim 1, characterized in that, The feature extraction model includes a powder feature extraction branch and a fiber feature extraction branch. The feeding feature vector includes a powder feeding feature vector and a fiber feeding feature vector. The process of generating the feeding feature vector through short-term feature extraction using the feature extraction model includes: The actual feeding speed of the flame retardant powder, the target feeding speed of the powder, and the actual operating data of the feeder are used as the powder branch input data. The powder branch input data is then passed through the powder multi-scale convolutional layer of the powder feature extraction branch to generate multi-scale powder feeding features. The multi-scale powder feeding features and the powder branch input data are weighted by multi-scale feature guidance through the powder attention mechanism of the powder feature extraction branch to generate the powder feeding feature vector. The actual feeding speed of the glass fiber, the target feeding speed of the fiber, and the actual operating data of the feeder are used as fiber branch input data. The fiber branch input data are then processed through the fiber feature extraction branch multi-scale convolutional layer to generate multi-scale fiber feeding features. The multi-scale fiber feeding features and fiber branch input data are weighted by multi-scale feature guidance through the fiber attention mechanism of the fiber feature extraction branch to generate the fiber feeding feature vector.

3. The method for preparing environmentally friendly high-heat-wire, high-impact PBT particles according to claim 2, characterized in that, The multi-scale powder feeding characteristics include powder fluctuation characteristics, powder bridging trend characteristics, and powder flow characteristics. The process of generating the powder feeding feature vector through multi-scale convolutional layers and powder attention mechanisms includes: The powder branch input data is passed through the first powder convolutional layer, the second powder convolutional layer and the third powder convolutional layer of the powder multi-scale convolutional layer respectively to generate powder fluctuation characteristics, powder bridging trend characteristics and powder flow characteristics in sequence; The standard deviation, skewness, and kurtosis of the powder branch input data are respectively processed through the powder attention mechanism to generate fluctuation attention weight, agglomeration attention weight, and bridging attention weight. The fluctuation attention weight, agglomeration attention weight, and bridging attention weight are then weighted and summed with the powder fluctuation characteristics, powder bridging trend characteristics, and powder flow characteristics to generate the powder feeding feature vector. The multi-scale fiber feeding features include fiber breakage features, fiber entanglement trend features, and fiber transport features. The process of generating fiber feeding feature vectors through fiber multi-scale convolutional layers and fiber attention mechanisms includes: The fiber branch input data is passed through the first fiber convolutional layer, the second fiber convolutional layer, and the third fiber convolutional layer of the fiber multi-scale convolutional layer respectively to generate fiber breakage characteristics, fiber entanglement trend characteristics, and fiber transport characteristics in sequence. The mean, variance, and information entropy of the fiber branch input data are respectively processed through the fiber attention mechanism to generate breakage attention weight, entanglement attention weight, and feeding attention weight. The breakage attention weight, entanglement attention weight, and feeding attention weight are then weighted and summed with the fiber breakage feature, fiber entanglement trend feature, and fiber transport feature to generate the fiber feeding feature vector. The convolution kernel sizes of the convolutional layers, from smallest to largest, are as follows: the first powder convolutional layer, the first fiber convolutional layer, the second powder convolutional layer, the second fiber convolutional layer, the third powder convolutional layer, and the third fiber convolutional layer.

4. The method for preparing environmentally friendly high-heat-wire, high-impact PBT particles according to claim 2, characterized in that, The LSTM model includes a fast LSTM branch for powders, a slow LSTM branch for fibers, and an interaction layer. The process of generating feed bridging and agglomeration features using the LSTM model includes: The powder feeding feature vector and its change are respectively passed through a fast LSTM branch for powder with fiber interaction input gate and rate of change enhancement forget gate to generate the powder hidden state; The fiber feeding feature vector and its changes are passed through a slow LSTM branch with a trend-enhancing input gate and a smooth forgetting gate to generate the fiber hidden state. The powder hidden state and fiber hidden state are passed through an interaction layer to generate the feed bridging agglomeration feature.

5. The method for preparing environmentally friendly high-heat-wire, high-impact PBT particles according to claim 4, characterized in that, The process of generating a hidden state for powder includes: The powder hidden state, the powder feeding feature vector, and the spliced ​​vector of the fiber hidden state from the previous time step are linearly mapped to generate the fiber interaction input gate and the initial forget gate. The differences of the powder feeding feature vectors from multiple time steps are weighted and summed on the initial forget gate to generate the rate of change enhanced forget gate. The powder hidden state at the current time step is updated based on the fiber interaction input gate and the rate of change enhanced forget gate. The process of generating the hidden state of fibers includes: The glass fiber feeding trend value is calculated based on the difference between the actual feeding speed and the target feeding speed of the glass fiber within a set time window. The initial input gate based on the fiber feeding feature vector and the glass fiber feeding trend value are weighted and summed to generate the trend enhancement input gate. The initial forgetting gate based on the fiber feeding feature vector is weighted and summed with the rate of change of the fiber feeding feature vector to generate the smooth forgetting gate. The hidden state of the fiber at the current time step is updated based on the trend enhancement input gate and the smooth forgetting gate.

6. The method for preparing environmentally friendly high-heat-wire, high-impact PBT particles according to claim 4, characterized in that, The process of generating feed bridging cluster features through the interaction layer includes: The powder hidden state and the fiber hidden state are weighted and summed to generate a linear term; The interaction term is obtained by multiplying the powder hidden state and the fiber hidden state element by element. The linear term and the interaction term are weighted and summed to generate a comprehensive hidden state, and the comprehensive hidden state, the powder hidden state, and the fiber hidden state at the final time step are used as the feed bridging and agglomeration features.

7. The method for preparing environmentally friendly high-heat-wire, high-impact PBT particles according to claim 2, characterized in that, The process of adjusting the desired powder feed speed and the desired fiber feed speed includes: Based on the difference between the predicted powder feeding speed and the target powder feeding speed, a powder adjustment amount is generated. Based on the powder adjustment amount and the calculation of the current powder feeding speed, a desired powder feeding speed is generated. The fiber adjustment amount is generated based on the difference between the predicted fiber feeding speed and the target fiber feeding speed, and the desired fiber feeding speed is generated based on the calculation of the fiber adjustment amount and the current fiber feeding speed.

8. The method for preparing environmentally friendly high-heat-wire high-impact PBT particles according to any one of claims 1 to 7, characterized in that, Also includes: Based on the mean square error between the predicted powder feeding rate and the optimal powder feeding rate of the sample, and the mean square error between the predicted fiber feeding rate and the optimal fiber feeding rate of the sample, a main loss term is constructed. A flow stability loss term is constructed based on the ratio of the predicted powder feeding rate to the predicted fiber feeding rate and the ratio of the theoretical powder flow rate to the theoretical fiber flow rate. A comprehensive loss function is constructed based on the weighted sum of the main loss term and the traffic stability loss term. The feature extraction model, the LSTM model, and the fully connected network are then collaboratively optimized and trained using the comprehensive loss function.

9. An environmentally friendly high-glow wire, high-impact PBT particle prepared by the method described in any one of claims 1 to 8, characterized in that, By weight, it comprises 30-50 parts PBT resin, 5-40 parts glass fiber and 23.9-60.4 parts composite flame retardant, wherein the composite flame retardant includes halogenated flame retardant, nitrogen-based MCA flame retardant and auxiliary additives. The halogenated flame retardant, by weight, comprises 8-15 parts brominated epoxy, 5-15 parts brominated polystyrene and 1-5 parts antimony trioxide; The nitrogen-based MCA flame retardant comprises, by weight, 3-10 parts nitrogen-based flame retardant MCA, 2-6 parts stannate and 1-3 parts silicate.

10. The environmentally friendly high-heat-wire, high-impact PBT granules according to claim 9, characterized in that, The auxiliary additives, by weight, include 3-5 parts toughening agent, 0.2-0.4 parts antioxidant, 0.5 parts dispersant and 0.2-0.5 parts coupling agent.

Images