Method and system for managing perfluorotributylamine product life cycle

Abstract

The application discloses a perfluorotributylamine product life cycle management method and system, the method comprises the following steps: obtaining the whole life cycle stage sequence of perfluorotributylamine product from raw material input to final disposal; collecting material balance data, equipment working condition data and environmental emission data for each life cycle stage, and inputting the material balance data into a recursive convolution deep fusion network to perform cross-stage material flow tracking processing, and generating an inter-stage material migration characteristic map; calling a pre-trained life cycle impact assessment model to perform multi-objective environmental impact decoupling processing on the inter-stage material migration characteristic map, and outputting an environmental impact contribution degree vector corresponding to each life cycle stage; performing reverse source attribution processing on the whole life cycle stage sequence based on the environmental impact contribution degree vector, marking at least one key control stage with a contribution degree exceeding a dynamic threshold, and outputting the key control stage as a process optimization anchor point.

Description

Technical Field

[0001] The present invention relates to the technical field of product life cycle management, and specifically to a method and system for product life cycle management based on perfluorotributylamine. Background Technique

[0002] As a high-value-added fluorinated fine chemical, perfluorotributylamine is widely used in fields such as electronic cooling and pharmaceutical intermediates. Its production process covers multiple stages including fluorite mining, electrolytic fluorination, catalytic fluorination, purification and rectification, and waste disposal. The material flow is tightly coupled, and the components of environmental emissions are complex. Existing life cycle management methods mostly rely on static inventory analysis, and use linear superposition or fixed distribution coefficients to process the material transfer relationships between different stages, making it difficult to reflect the dynamic migration behavior of materials between different stages and the hidden dependency patterns in the actual production process. At the same time, environmental impact assessment usually converts multiple environmental impact categories into a single comprehensive index or only conducts simple classification weighting, and cannot decouple the stage attribution of multiple types of environmental impacts such as greenhouse effect, acidification effect, and ecological toxicity under a unified framework, resulting in the assessment result having too coarse granularity to accurately identify the process links that contribute the most to specific environmental goals.

[0003] In terms of cross-stage material flow tracking, traditional methods perform material balance calculations with stages as independent units, ignoring the chain effects of fluctuations in previous operating conditions and deviations in reaction conversion rates on the input materials of subsequent stages, blurring the continuity of the material migration path and the transfer losses between stages, and making it difficult to form a high-resolution material migration feature expression that can be utilized by deep learning models. In terms of decoupling and attributing multi-objective environmental impacts, existing attribution techniques mostly adopt the method of directly apportioning forward from the total emissions according to the production ratio, lacking a reverse layer-by-layer propagation mechanism based on the material migration coefficient matrix between stages, and the threshold setting uses fixed empirical values, unable to adaptively adjust according to the actual contribution degree distribution of each batch of products, resulting in inaccurate marking of key control stages and a lack of pertinence in the process optimization direction. Summary of the Invention

[0004] The present invention provides a method and system for product life cycle management based on perfluorotributylamine, which realizes the refined tracking and characterization of cross-stage material flow through a recursive convolutional deep fusion network, and uses a multi-head attention mechanism to decouple and reverse trace the attribution of multi-objective environmental impacts, and marks the key process control stages with an adaptive dynamic threshold, providing an anchor point for the precise process optimization of perfluorotributylamine products.

[0005] The object of the present invention can be achieved through the following technical solutions:

[0006] The present invention provides a method for product life cycle management based on perfluorotributylamine, and the method includes:

[0007] Obtain the full life cycle sequence of perfluorotributylamine products from raw material input to final disposal. The full life cycle sequence includes the ore pretreatment stage, the electrolytic hydrofluoric acid production stage, the catalytic fluorination synthesis stage, the purification and distillation stage, the product application stage, and the waste disposal stage.

[0008] Material balance data, equipment operating data, and environmental emission data are collected for each life cycle stage. The material balance data is then input into a recursive convolutional deep fusion network to perform cross-stage material flow tracking processing, generating inter-stage material migration feature maps. Local material migration patterns are captured through spatial convolutional layers in the recursive convolutional deep fusion network, and cross-stage dependencies are encoded using long short-term memory layers, effectively improving the representation accuracy of material flow evolution.

[0009] The pre-trained life cycle impact assessment model is invoked to perform multi-objective environmental impact decoupling processing on the inter-stage material migration feature map, and output the environmental impact contribution vector corresponding to each life cycle stage. By using the multi-head attention decomposition layer in the model to focus on multiple environmental indicators such as greenhouse effect, acidification effect and ecotoxicity, fine-grained decoupling of impact contribution is achieved, making the stage-level environmental impact assessment more objective.

[0010] Based on the environmental impact contribution vector, reverse attribution processing is performed on the full life cycle stage sequence to mark at least one key control stage whose contribution exceeds the dynamic threshold. The key control stage is then output as the process optimization anchor point, thereby distributing the environmental impact pressure in reverse along the material migration path and accurately identifying the weak links that need to be prioritized for regulation in the entire chain.

[0011] As a technical solution of the present invention, the acquisition of the full life cycle sequence of perfluorotributylamine products from raw material input to final disposal includes: collecting the mining batch identifier and transportation route information of fluorite raw materials; allocating the fluorite raw materials to the input queue of the ore pretreatment stage according to the mining batch identifier; performing state machine modeling on the start-up and shutdown events of the electrolyzer in the electrolysis to produce hydrofluoric acid stage, generating an electrolysis process state transition sequence, and aligning and binding the electrolysis process state transition sequence with the feed sequence of the catalytic fluorination synthesis stage to ensure... Strict coupling of material flow and time dimension; component identification and tracking processing of the top and bottom effluent streams of the multi-stage distillation column in the purification and distillation stage, generating component flow direction tags, and attaching the component flow direction tags to the input material descriptor in the product application stage; timestamp alignment processing of the operation logs of the waste gas incineration unit, wastewater neutralization unit and solid waste landfill unit in the waste disposal stage, establishing an association index between the aligned operation logs and the output residual streams of the product application stage, thereby forming an end-to-end traceable data chain.

[0012] Preferably, the step of collecting material balance data, equipment operating condition data, and environmental emission data for each life cycle stage, and inputting the material balance data into a recursive convolutional deep fusion network to perform cross-stage material flow tracking processing and generate an inter-stage material migration feature map, includes: collecting the difference in the inflow and outflow quality of fluorite tailings as first material balance data for the ore pretreatment stage; collecting the ratio of anode gas emissions to cathode liquid production as second material balance data for the electrolytic hydrofluoric acid production stage; and collecting the ratio of perfluorotributylamine crude product production to catalyst consumption as third material balance data for the catalytic fluorination synthesis stage; concatenating the first, second, and third material balance data into a multi-channel material flow tensor in stage time sequence; inputting the multi-channel material flow tensor into the spatial convolutional layer of the recursive convolutional deep fusion network for local material migration pattern extraction processing to generate a local migration feature map; and inputting the local migration feature map into the long short-term memory layer of the recursive convolutional deep fusion network for cross-stage dependency encoding processing to generate the inter-stage material migration feature map. In this process, the local migration feature map corresponding to the i-th life cycle stage is input into the forgetting gate unit of the Long Short-Term Memory (LSTM) layer. The forgetting gate unit reads the hidden state vector of the previous stage and the input feature vector of the current stage, and outputs an inter-stage material discarding weight matrix. The local migration feature map is then input into the input gate unit of the LSM layer. The input gate unit calculates candidate memory unit vectors and updates the memory unit state of the current stage according to the inter-stage material discarding weight matrix. The updated memory unit state is then input into the output gate unit of the LSM layer. The output gate unit generates the hidden state vector of the current stage as the feature representation of the corresponding stage in the inter-stage material migration feature map. This approach preserves the local material change patterns of each stage while capturing the long-range influence relationships between upstream and downstream stages.

[0013] Furthermore, the step of calling the pre-trained life cycle impact assessment model to perform multi-objective environmental impact decoupling processing on the inter-stage material migration feature map and outputting the environmental impact contribution vector corresponding to each life cycle stage includes: inputting the inter-stage material migration feature map into the global average pooling layer of the life cycle impact assessment model for feature dimensionality reduction processing to generate a stage aggregated feature vector; inputting the stage aggregated feature vector into the multi-head attention decomposition layer of the life cycle impact assessment model, wherein the multi-head attention decomposition layer includes a greenhouse effect attention head, an acidification effect attention head, and an ecotoxicity attention head, and each attention head independently calculates the impact weight distribution of each life cycle stage; and performing vector concatenation processing on the first impact weight vector output by the greenhouse effect attention head, the second impact weight vector output by the acidification effect attention head, and the third impact weight vector output by the ecotoxicity attention head to generate the environmental impact contribution vector. Specifically, the greenhouse effect attention head generates a greenhouse effect query vector, a greenhouse effect key vector, and a greenhouse effect value vector by performing matrix multiplication operations on the stage aggregated feature vector with the first query weight matrix, the first key weight matrix, and the first value weight matrix, respectively. It then calculates the dot product similarity matrix between the greenhouse effect query vector and the greenhouse effect key vector, and performs inter-stage weight normalization processing via a flexible maximum value normalization unit to generate the first influence weight vector. Finally, it performs a weighted summation operation on the greenhouse effect value vector and the first influence weight vector to generate a greenhouse effect weighted feature vector, which is then passed to the output splicing layer. Other attention heads independently generate influence weights for their respective categories in a similar manner, achieving decoupled expression of the contribution of a single stage to different environmental indicators and avoiding mutual interference between different influence categories.

[0014] In a preferred embodiment of the present invention, reverse attribution processing is performed on the entire life cycle stage sequence based on the environmental impact contribution vector to mark at least one key control stage whose contribution exceeds a dynamic threshold, and the key control stage is output as a process optimization anchor point. This includes: performing a reverse traversal scan from the product application stage to the ore pretreatment stage on each component of the environmental impact contribution vector; reversing the environmental impact contribution of the current stage to the input material nodes of the preceding stage according to the material migration path; calculating the dynamic threshold of each life cycle stage based on the cumulative contribution value after reverse allocation, wherein the dynamic threshold is twice the average contribution value of the entire life cycle stage sequence; marking the life cycle stage whose contribution exceeds the dynamic threshold as the key control stage, and extracting the equipment identifier and operating parameter range corresponding to the key control stage as the process optimization anchor point. During reverse allocation, the material migration coefficient matrix between stage j and stage j minus one in the full life cycle stage sequence is obtained. The row index of the material migration coefficient matrix corresponds to the output material type of stage j, and the column index corresponds to the input material type of stage j minus one. The environmental impact contribution vector of stage j is multiplied by the material migration coefficient matrix in reverse to generate the contribution vector to be allocated for stage j minus one. The contribution vector to be allocated is accumulated with the original direct environmental impact contribution vector of stage j minus one to update the cumulative contribution vector of stage j minus one. The above steps are repeated until the ore pretreatment stage is reached to generate the final reverse cumulative contribution for each life cycle stage. This attribution mechanism can truly reflect the potential environmental impact of upstream processes on downstream processes and improve the reliability of key node identification.

[0015] In some technical solutions of the present invention, the material balance data, equipment operating condition data and environmental emission data are standardized before being input into the recursive convolutional deep fusion network. The standardization process uses the Z-score normalization method to adjust each data dimension to a zero-mean unit variance distribution, thereby eliminating the impact of dimensional differences on the stability of network training.

[0016] As a further improvement of the present invention, the method further includes a process parameter optimization process performed after the key control stages are marked. This process parameter optimization process includes: extracting a set of historical operating parameter records corresponding to the key control stages, the set of historical operating parameter records including electrolysis temperature records, distillation reflux ratio records, and catalyst addition rate records; associating and aligning the set of historical operating parameter records with the feature vectors of the corresponding stages in the inter-stage mass migration feature map to generate parameter influence feature pairs; calling a Bayesian optimization surrogate model to perform Gaussian process regression fitting on the parameter influence feature pairs, outputting a Pareto front set of process parameters, and using the parameter combination closest to the ideal point in the Pareto front set of process parameters as the recommended optimization parameters. Through this method, actionable process improvement suggestions can be provided while taking into account multiple environmental indicators, avoiding the transfer of environmental burden caused by optimization based solely on a single objective.

[0017] As another technical solution of the present invention, the method further includes a lifecycle inventory dynamic update process performed after outputting the process optimization anchor point. The lifecycle inventory dynamic update process includes: acquiring measured emission data after adjusting process parameters during the key control stage; comparing the measured emission data with preset baseline emission data in the process optimization anchor point to generate an emission deviation vector; inputting the emission deviation vector into a preset inventory correction mapping layer, whereby the inventory correction mapping layer determines the correction direction based on the sign direction of the emission deviation vector and determines the correction step size based on the absolute value of the emission deviation vector; updating the inventory factor values ​​for all stages in the full lifecycle stage sequence from the key control stage to the waste disposal stage according to the correction direction and correction step size, generating an updated lifecycle inventory database, and feeding the updated lifecycle inventory database back to the input layer of the recursive convolutional deep fusion network for inter-stage material flow tracking processing of the next batch of products. This forms a closed-loop management mechanism of "evaluation-optimization-verification-update," continuously improving the timeliness and accuracy of the lifecycle inventory data.

[0018] This invention also includes a perfluorotributylamine (PFTB) product lifecycle management system. This system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the aforementioned perfluorotributylamine (PFTB) product lifecycle management method. By integrating technologies such as deep network material flow tracing, multi-objective environmental impact decoupling, and reverse attribution, the system can automatically complete the entire process analysis from data acquisition to optimization decision-making, providing efficient and accurate technical support for green process improvement and full lifecycle environmental management of perfluorotributylamine products.

[0019] The beneficial effects of this invention are:

[0020] Material balance data collected at each stage of the perfluorotributylamine product's lifecycle are concatenated into a multi-channel mass flow tensor in chronological order and input into a recursive convolutional deep fusion network. Spatial convolutional layers perform local mass migration pattern extraction on the multi-channel mass flow tensor, generating local migration feature maps to capture the transformation relationships between input materials and outputs within the same stage. Long Short-Term Memory (LSTM) layers encode cross-stage dependencies in the local migration feature maps. A forgetting gating unit generates an inter-stage mass discard weight matrix based on the hidden state vector of the previous stage and the input feature vector of the current stage. The input gating unit calculates candidate memory unit vectors and updates the memory unit state of the current stage. The output gating unit generates the hidden state vector of the current stage as the feature representation of the corresponding stage in the inter-stage mass migration feature map. This structure allows the amount of material retained, lost, and transferred downstream between stages to be preserved and transferred across stages in the form of memory unit states. The resulting inter-stage material migration feature map can characterize the nonlinear transfer law and delayed response characteristics of material flows such as fluorite tailings, anode gas escape, and catalyst consumption between stages such as ore pretreatment, electrolytic hydrofluoric acid production, catalytic fluorination synthesis, and purification distillation. It overcomes the shortcomings of traditional independent material balance that cannot connect the dynamic correlation of material flows between stages.

[0021] The inter-stage material migration feature map is input into the global average pooling layer of the life cycle impact assessment model for feature dimensionality reduction, generating stage aggregated feature vectors, which are then input into a multi-head attention decomposition layer. This multi-head attention decomposition layer includes attention heads for greenhouse effect, acidification effect, and ecotoxicity. Each attention head independently calculates the impact weight distribution for each life cycle stage, generating a weighted feature vector corresponding to the environmental impact category. These vectors are then concatenated to generate an environmental impact contribution vector. Based on this contribution vector, a reverse traversal scan of the entire life cycle stage sequence is performed to obtain the material migration coefficient matrix between adjacent stages. The environmental impact contribution vector of the current stage is multiplied in reverse by the material migration coefficient matrix to generate the unassigned contribution vector for the preceding stage. This vector is then accumulated with the original direct environmental impact contribution vector of the preceding stage, propagating stage by stage until the ore pretreatment stage, generating the final reverse cumulative contribution for each life cycle stage. A dynamic threshold is calculated based on the cumulative contribution of each stage in the entire life cycle stage sequence. Life cycle stages with contributions exceeding the dynamic threshold are marked as critical control stages, and the corresponding equipment identifiers and operating parameter ranges are extracted as process optimization anchor points. The multi-attention decomposition mechanism enables independent decoupling of the attribution of different categories of environmental impacts between stages, avoiding mutual interference between different environmental indicators. During backpropagation, a mass migration coefficient matrix is ​​used for inverse allocation, ensuring that the attribution of environmental impacts is strictly transmitted in reverse along the actual mass flow direction. This allows for precise identification of process stages that significantly contribute to individual indicators such as the greenhouse effect, acidification effect, or ecotoxicity. Dynamic thresholds are adaptively determined based on the actual contribution distribution of each batch, avoiding the problem of poor applicability of fixed thresholds across different product batches. Attached Figure Description

[0022] The invention will now be further described with reference to the accompanying drawings.

[0023] Figure 1 This is a flowchart based on the product lifecycle management method for perfluorotributylamine;

[0024] Figure 2 This is a flowchart of the recursive convolution deep fusion processing of perfluorotributylamine throughout its entire life cycle material balance.

[0025] Figure 3 This is a flowchart for the reverse attribution process of environmental impact contribution. Detailed Implementation

[0026] 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, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] See Figure 1 This invention provides a lifecycle management method for perfluorotributylamine (PFTB) products, comprising: acquiring a full lifecycle sequence of PFTB products from raw material input to final disposal, wherein the full lifecycle sequence includes an ore pretreatment stage, an electrolytic hydrofluoric acid production stage, a catalytic fluorination synthesis stage, a purification and distillation stage, a product application stage, and a waste disposal stage; collecting material balance data, equipment operating data, and environmental emission data for each lifecycle stage, and inputting the material balance data into a recursive convolutional deep fusion network to perform cross-stage material flow tracing processing to generate an inter-stage material migration feature map; calling a pre-trained lifecycle impact assessment model to perform multi-objective environmental impact decoupling processing on the inter-stage material migration feature map, and outputting an environmental impact contribution vector corresponding to each lifecycle stage; performing reverse source attribution processing on the full lifecycle sequence based on the environmental impact contribution vector, marking at least one key control stage whose contribution exceeds a dynamic threshold, and outputting the key control stage as a process optimization anchor point.

[0028] In specific implementation, please refer to Figure 2 The process of obtaining the complete lifecycle sequence of perfluorotributylamine (PFTB) products from raw material input to final disposal includes: collecting the mining batch identifier and transportation route information of fluorite raw materials; and allocating the fluorite raw materials to the input queue of the ore pretreatment stage according to the mining batch identifier. The mining batch identifier is recorded by laser-engraving a QR code on the fluorite raw material packaging unit, and the transportation route information is uploaded to the central database by the vehicle-mounted GPS terminal at fixed time intervals. The input queue of the ore pretreatment stage is a first-in, first-out (FIFO) queue, and each element in the queue includes the mining batch identifier, the entry timestamp, and the estimated quality of the fluorite raw material. The fluorite raw materials enter the ore pretreatment stage sequentially according to the order of the mining batch identifiers.

[0029] The hydrofluoric acid electrolysis stage comprises multiple electrolytic cells. State machine modeling is used to handle the start-up and shutdown events of these cells. A finite set of states is defined for each cell, including shutdown, preheating, normal electrolysis, pause, and fault states. Current, voltage, and temperature signals from the electrolytic cells are collected. State switching conditions are determined based on current and voltage thresholds, generating a state transition sequence bound to the cell identifier. Each entry in the state transition sequence includes a state name, a timestamp of entering the state, and the duration of the state. The state transition sequence is aligned with the feed sequence of the catalytic fluorination synthesis stage: continuous periods in the normal electrolysis state are extracted from the state transition sequence, and the hydrofluoric acid production rate curve is calculated. The time axis of the hydrofluoric acid production rate curve is dynamically time-aligned with the cumulative feed rate curve recorded by the hydrofluoric acid feed flow meter in the catalytic fluorination synthesis stage. The aligned time index relationship is stored as a binding mapping table.

[0030] The purification and distillation stage comprises three distillation columns. Component identification and tracking are performed on the top and bottom feed streams from each column. Online gas chromatographs are installed at the top and bottom feed outlets of each column. The feed streams are sampled and analyzed at fixed intervals to obtain the mass fractions of perfluorotributylamine (PFMA), isomers, and fluorine-containing impurities. The analysis results are combined with the sampling timestamp, distillation column identifier, and feed location to generate component flow direction tags. These tags are appended to the input material descriptor in the product application stage. The input material descriptor is a JSON data structure containing the material batch number, a list of component flow direction tags, and a production timestamp. The product application stage receives the input material descriptor and determines the material's compliance based on the PFMA mass fraction in the component flow direction tags.

[0031] The waste disposal phase includes waste gas incineration, wastewater neutralization, and solid waste landfill. The operation logs of these three units are timestamped for alignment. The waste gas incineration log includes incineration temperature and flue gas flow rate records; the wastewater neutralization log includes influent pH and chemical dosage records; and the solid waste landfill log includes landfill material weighing records. All three logs are aligned using UTC timestamps, and nearest-neighbor interpolation is used to resample log sequences from different sampling frequencies to a unified time grid. An association index is established between the aligned operation logs and the residual flow streams from the product application phase. The residual flow streams from the product application phase include the waste perfluorotributylamine container identifier, residual liquid mass, and discharge timestamp. The corresponding time window's operation record for the waste disposal unit is retrieved from the aligned operation logs using the discharge timestamp, generating an association index entry and storing it in the association database.

[0032] Material balance data, equipment operating condition data, and environmental emission data are collected for each stage of the life cycle. The material balance data collection process includes: for the ore pretreatment stage, the difference between the incoming and outgoing mass of fluorite tailings is collected as the first material balance data. The incoming mass is the difference between the total mass of the fluorite raw material transport vehicles entering the plant and the empty vehicle mass, as measured by the weighbridge. The outgoing mass is the difference between the total mass of the tailings transport vehicles leaving the plant and the empty vehicle mass, as measured by the weighbridge. The first material balance data is the difference between the incoming and outgoing mass. For the hydrofluoric acid electrolysis stage, the ratio of anode gas emissions to cathode liquid production is collected as the second material balance data. The anode gas emissions are cumulatively measured by a thermal gas flow meter on the anode gas collection pipeline, and the cathode liquid production is cumulatively measured by a Coriolis mass flow meter on the cathode liquid collection pipeline. The second material balance data is the ratio of the cumulative anode gas emissions to the cumulative cathode liquid production. The ratio of crude perfluorotributylamine (PFTB) produced to catalyst consumed during the catalytic fluorination synthesis stage was used as the third material balance data. The amount of crude PFTB produced was calculated by an online gas chromatograph at the outlet of the synthesis reactor combined with a flow meter. The amount of catalyst consumed was obtained by the mass difference of the catalyst storage tank weighing sensor during the reaction cycle. The third material balance data is the ratio of crude PFTB produced to catalyst consumed.

[0033] The first, second, and third material balance data are concatenated in a phase-sequential order to form a multi-channel material flow tensor. The first dimension of the multi-channel material flow tensor corresponds to the phase index of the entire lifecycle phase sequence, and the second dimension corresponds to the material balance data channel. The multi-channel material flow tensor is then input into the spatial convolutional layer of a recursive convolutional deep fusion network for local material migration pattern extraction, generating a local migration feature map. The spatial convolutional layer consists of multiple one-dimensional convolutional kernels that slide along the phase index direction. Each one-dimensional convolutional kernel performs convolution operations with local regions of the multi-channel material flow tensor across several consecutive phases, outputting each channel of the local migration feature map. A linear rectified activation function is then applied after the spatial convolutional layer.

[0034] The local migration feature map is input into the Long Short-Term Memory (LSTM) layer of a recursive convolutional deep fusion network for cross-stage dependency encoding, generating an inter-stage material migration feature map. The LSM layer contains a forgetting gate unit, an input gate unit, and an output gate unit. In practice, the local migration feature map corresponding to the i-th lifecycle stage is input into the forgetting gate unit of the LSM layer. The forgetting gate unit reads the hidden state vector of the previous stage and the input feature vector of the current stage, and outputs the inter-stage material discarding weight matrix. The forward propagation formula for the forgetting gate unit is:

[0035]

[0036] in, This represents the inter-stage material discarding weight matrix for the i-th life cycle stage. This represents the sigmoid activation function. This represents the forget gate weight matrix. This represents the hidden state vector for the (i-1)th lifecycle stage. This represents the input feature vector obtained by flattening the local transfer feature map corresponding to the i-th lifecycle stage. Represents the hidden state vector With input feature vector Vector concatenation, This represents the forget gate bias vector. The forget gate weight matrix. The dimension is Forget gate bias vector The dimension is ,in Let be the dimension of the hidden state vector. The dimension of the input feature vector. Forget gate weight matrix. and forget gate bias vector The value of is obtained through supervised pre-training of a recursive convolutional deep fusion network on a historical full life cycle material balance dataset. The training objective is to minimize the weighted sum of the inter-stage material flow reconstruction error and the next stage environmental impact prediction error.

[0037] The local migration feature map is input to the input gating unit of the Long Short-Term Memory (LSTM) layer. The input gating unit calculates the candidate memory cell vector and updates the memory cell state of the current stage according to the inter-stage material discarding weight matrix. The updated memory cell state is input to the output gating unit of the LSM layer. The output gating unit generates the hidden state vector of the current stage as the feature representation of the corresponding stage in the inter-stage material migration feature map. The above process is repeated to traverse all stages in the entire life cycle stage sequence, and the hidden state vectors corresponding to all stages are arranged in stage order to form the inter-stage material migration feature map.

[0038] In specific implementation, the process of decoupling the multi-objective environmental impact of the inter-stage material migration feature map by calling the pre-trained life cycle impact assessment model includes: inputting the inter-stage material migration feature map into the global average pooling layer of the life cycle impact assessment model for feature dimensionality reduction to generate stage aggregated feature vectors; inputting the stage aggregated feature vectors into the multi-head attention decomposition layer of the life cycle impact assessment model, which includes a greenhouse effect attention head, an acidification effect attention head, and an ecotoxicity attention head, with each attention head independently calculating the impact weight distribution of each life cycle stage; and concatenating the first impact weight vector output by the greenhouse effect attention head, the second impact weight vector output by the acidification effect attention head, and the third impact weight vector output by the ecotoxicity attention head to generate an environmental impact contribution vector.

[0039] The lifecycle impact assessment model consists of a global average pooling layer, a multi-head attention decomposition layer, and an output concatenation layer. The global average pooling layer receives inter-stage mass migration feature maps as input, and the dimensions of the inter-stage mass migration feature maps are S rows, ... Column S represents the total number of stages in the entire lifecycle stage sequence. This represents the dimension of the feature representation vector for each stage. The global average pooling layer performs an arithmetic average operation on the inter-stage material migration feature maps along the stage dimension, and then calculates the S values. 3D feature vectors compressed into one The stage-aggregated feature vector is dimensional. The output of the global average pooling layer is connected to the input of the multi-head attention decomposition layer, and the output of the multi-head attention decomposition layer is connected to the input of the output concatenation layer.

[0040] The multi-head attention decomposition layer contains three parallel attention head modules: a greenhouse effect attention head, an acidification effect attention head, and an ecotoxicity attention head. Each of the three attention head modules receives the same aggregated feature vector for a given stage and independently outputs an influence weight vector. The greenhouse effect attention head outputs a first influence weight vector with dimension S; the acidification effect attention head outputs a second influence weight vector with dimension S; and the ecotoxicity attention head outputs a third influence weight vector with dimension S. The output concatenation layer concatenates the first, second, and third influence weight vectors according to their dimensions to generate a 3S-dimensional environmental impact contribution vector.

[0041] Within the greenhouse effect attention head, the stage-aggregated feature vector is multiplied by the first query weight matrix of the greenhouse effect attention head to generate the greenhouse effect query vector. The dimension of the first query weight matrix is... OK, List, The dimension of the query vector is set to 64. A matrix multiplication operation is performed between the stage aggregated feature vector and the first-key weight matrix of the greenhouse effect attention head. The dimension of the first-key weight matrix is... Line, (S×) )List, The dimension of the key vector is set to 64. The result of the multiplication operation is a one-dimensional vector, which is then rearranged into S rows. The greenhouse effect key matrix is ​​a columnar structure, where each row corresponds to a key vector for a lifecycle stage. The stage-aggregated feature vectors are then multiplied by the first-value weight matrix of the greenhouse effect attention head. The first-value weight matrix has dimensions of... Line, (S×) )List, The dimension of the value vector is set to 64. The results of the multiplication operation are rearranged into S rows. The greenhouse effect value matrix consists of columns, with each row of the greenhouse effect value matrix corresponding to a value vector of a life cycle stage.

[0042] Calculate the dot product similarity between the greenhouse effect query vector and the key vectors in each row of the greenhouse effect key matrix, generating a similarity vector of dimension S. Input this similarity vector into the flexible maximum normalization unit of the greenhouse effect attention head for inter-stage weight normalization, generating the first influence weight vector.

[0043]

[0044] in, This represents the influence weight value corresponding to the s-th life cycle stage in the first influence weight vector. The value of is an integer from 1 to S; The dimension is The greenhouse effect query vector; Represents the greenhouse effect bond matrix. The key vector corresponding to the row, To sum the index variable, iterate through all stages from 1 to S; This represents the dimension of the key vector, with a value of 64. This represents an exponential function with base e to the natural constant. Each value in the similarity vector is divided by... Then, perform exponential operations, and divide by the sum of all exponentially calculated values ​​to obtain the normalized weight values. The weight values ​​constitute the first influence weight vector.

[0045] The greenhouse effect value matrix and the first impact weight vector are weighted and summed to generate a greenhouse effect weighted feature vector, which is then passed to the output concatenation layer of the life cycle impact assessment model. The weighted summation is calculated as follows: the s-th weight in the first impact weight vector is... Multiply by the first value in the greenhouse effect matrix The value vectors of each row are then summed element-wise to obtain the greenhouse effect weighted feature vector. The dimension of the greenhouse effect weighted feature vector is... The output concatenation layer receives weighted feature vectors generated by three attention heads, but does not participate in the concatenation of the final environmental impact contribution vector. The final output consists only of the concatenation of the three impact weight vectors.

[0046] The internal structures of the acidification effect attention head and the ecotoxicity attention head are the same as those of the greenhouse effect attention head. Each has its own independent second query weight matrix, second key weight matrix, second value weight matrix, as well as a third query weight matrix, third key weight matrix, and third value weight matrix. The acidification effect attention head outputs the second influence weight vector according to the same process, while the ecotoxicity attention head outputs the third influence weight vector.

[0047] The pre-training process of the life cycle impact assessment model employs supervised learning. When constructing the training dataset, multiple batches of full life cycle sequence data were collected from historical perfluorotributylamine (PFAT) production records. For each batch, inter-stage material migration feature maps were generated using the same method as in real-time processing. Simultaneously, for each batch, the ISO 14040 life cycle assessment method was used to calculate the characteristic environmental impact contribution value for each life cycle stage under the greenhouse effect, acidification effect, and ecotoxicity categories. The contribution values ​​of the three categories were concatenated in stage order to form a supervision label vector with the same dimension as the model output, i.e., the true environmental impact contribution vector for each batch. The total number of training samples was no less than 5000 batches, and the dataset was divided into training, validation, and test sets in an 8:1:1 ratio.

[0048] During pre-training, inter-stage material migration feature maps are input into the life cycle impact assessment model. Forward propagation yields the predicted environmental impact contribution vector, and the mean squared error loss function between the predicted and actual environmental impact contribution vectors is calculated. An adaptive moment estimation optimizer is used to update the parameters of the global average pooling layer, the weight matrices and bias vectors of all attention heads in the multi-head attention decomposition layer, and the parameters of the output concatenation layer. The initial learning rate of the adaptive moment estimation optimizer is set to 0.0001, the first-order moment decay coefficient is set to 0.9, the second-order moment decay coefficient is set to 0.999, and the mini-batch size is 16. The training iterations are set to 300 epochs, with each epoch completing a full traversal on the training set and calculating the loss value on the validation set. When the validation set loss value no longer decreases within 10 consecutive epochs, training is terminated, and the model parameters with the minimum validation set loss value are saved as the parameters for the pre-trained life cycle impact assessment model. After pre-training, the model's greenhouse effect attention head, acidification effect attention head, and ecotoxicity attention head learned the ability to decompose the contributions of different environmental impact categories from inter-stage material migration features. In the application stage, the model can be directly called to output the environmental impact contribution vector.

[0049] In practice, the process of performing reverse attribution processing on the entire life cycle stage sequence based on the environmental impact contribution vector includes: (See reference) Figure 3 For each component in the environmental impact contribution vector, a reverse traversal scan is performed from the product application stage to the ore pretreatment stage. The environmental impact contribution of the current stage is then reassigned to the input material nodes of the preceding stage according to the material migration path. The dynamic threshold for each life cycle stage is calculated based on the cumulative contribution value after reassignment. The dynamic threshold is twice the average contribution value of the entire life cycle stage sequence. Life cycle stages with contribution values ​​exceeding the dynamic threshold are marked as critical control stages, and the equipment identifiers and operating parameter ranges corresponding to the critical control stages are extracted as process optimization anchor points.

[0050] The environmental impact contribution vector is composed of impact components from three stages: greenhouse effect, acidification effect, and ecotoxicity. The reverse traversal scan starts at the three environmental impact category components corresponding to the product application stage in the environmental impact contribution vector and ends at the three environmental impact category components corresponding to the ore pretreatment stage.

[0051] During the reverse traversal scanning process, the material migration path between stage j and stage j minus one is modeled. The modeling method involves obtaining the material migration coefficient matrix between stage j and stage j minus one in the entire life cycle stage sequence. The row index of the material migration coefficient matrix corresponds to the output material type of stage j, and the column index corresponds to the input material type of stage j minus one. Each element in the material migration coefficient matrix represents the mass proportion of a certain output material from stage j to a certain input material in stage j minus one. The material migration coefficient matrix is ​​constructed as follows: Mass flow records of all output materials in stage j and all input materials in stage j minus one are extracted from the process flow diagram and material balance database of the perfluorotributylamine production process. For the u-th output material in stage j and the v-th input material in stage j minus one, the historical average mass proportion of the u-th output material in stage j flowing to the v-th input material in stage j minus one is calculated, and this historical average mass proportion is stored as the element value in the u-th row and v-th column of the material migration coefficient matrix. The sum of all element values ​​in each column of the mass migration coefficient matrix is ​​equal to 1, and the values ​​of all elements in the mass migration coefficient matrix are real numbers between 0 and 1.

[0052] In the j-th step of the reverse traversal scan, an inverse assignment operation is independently performed on each environmental impact category component in the environmental impact contribution vector. Taking the greenhouse effect category component as an example, the cumulative contribution vector of the j-th stage in the greenhouse effect category is multiplied inversely with the transpose of the mass migration coefficient matrix to generate the contribution vector to be assigned in the greenhouse effect category for the j-th minus one stage. The formula for the inverse matrix multiplication operation is:

[0053]

[0054] in, This represents the vector of unassigned contributions to the greenhouse effect category in the j-th minus one stage. The dimension is equal to the number of input material types in the j-th minus one stage. Each element value represents the contribution of the greenhouse effect from the j-th stage to the corresponding input material node in the j-minus-1 stage. This represents the mass migration coefficient matrix between stage j and stage j minus one. The dimension is OK, List, The output material types and quantities for stage j. For the input material type and quantity in the j-th minus one stage; superscript This represents the matrix transpose operation. The transposed mass transfer coefficient matrix has dimensions of 1. OK, List; This represents the cumulative contribution vector of stage j to the greenhouse effect category. The dimension is , Each element value represents the cumulative contribution of the greenhouse effect borne by the corresponding output material node in stage j.

[0055] After the reverse allocation in step j is completed, the contribution vector to be allocated will be... The cumulative contribution vector of the sub-j stage is updated by summing it with the original direct environmental impact contribution vector of the sub-j stage in the greenhouse effect category. The original direct environmental impact contribution vector of the sub-j stage is directly provided by the greenhouse effect component corresponding to the sub-j stage in the environmental impact contribution vector. The dimension of this vector is equal to the number of output material types in the sub-j stage. During the summation process, the contribution vector to be assigned is mapped to the output material nodes of the sub-j stage according to the correspondence between input and output materials. The mapping relationship is determined by the material flow relationship matrix within the sub-j stage. After mapping, it is added element-by-element to the direct environmental impact contribution vector to obtain the updated cumulative contribution vector of the greenhouse effect category for the sub-j stage. The acidification effect category component and the ecotoxicity category component are subjected to reverse allocation and cumulative update operations in the same way as the greenhouse effect category component.

[0056] The reverse traversal scan starts from the product application stage, repeatedly performing reverse allocation and cumulative update operations until the ore pretreatment stage is reached. After the traversal is completed, a complete final reverse cumulative contribution vector is generated for each lifecycle stage from the ore pretreatment stage to the product application stage. The final reverse cumulative contribution vector includes the cumulative contribution values ​​for the greenhouse effect category, the acidification effect category, and the ecotoxicity category.

[0057] The method for calculating the dynamic threshold for each life cycle stage based on the cumulative contribution value after reverse allocation is as follows: For each life cycle stage, the cumulative contribution values ​​of the three environmental impact categories in the final reverse cumulative contribution vector are summed to obtain the total cumulative contribution value for that life cycle stage; the arithmetic mean of the total cumulative contribution values ​​of all stages in the entire life cycle stage sequence is calculated; the arithmetic mean is multiplied by a coefficient of 2 to obtain the dynamic threshold. The basis for setting the coefficient of 2 as a fixed multiple is: in statistical process control, when the mean plus two standard deviations is used as the anomaly detection threshold, about 5% of the outliers can be marked for approximately symmetrically distributed data. Here, this multiple is used to construct the threshold baseline. Since the distribution of the total cumulative contribution value is affected by the scaling of the material migration coefficient matrix and presents an approximately normal distribution, using two times the mean as the dynamic threshold can adaptively separate high contribution stages without external prior standard deviation.

[0058] Lifecycle stages where the total cumulative contribution value exceeds a dynamic threshold are marked as critical control stages. The corresponding equipment identifiers and operating parameter ranges for these critical control stages are extracted as process optimization anchor points. Equipment identifiers are extracted from the equipment list of the corresponding stage in the full lifecycle stage sequence. The equipment list records the unique identification codes of the process equipment included in each lifecycle stage. Operating parameter ranges are extracted from the historical operating parameter record set of the critical control stage. The extraction method involves obtaining the 5th and 95th quantiles of all records in the historical operating parameter record set, using the 5th quantile as the lower limit of the operating parameter range and the 95th quantile as the upper limit. The marked critical control stage, the equipment identifier of the critical control stage, and the operating parameter range of the critical control stage are collectively encapsulated into a process optimization anchor point data structure and output to the process optimization module.

[0059] In practice, material balance data, equipment operating data, and environmental emission data undergo standardization before being input into the recursive convolutional deep fusion network. Standardization uses Z-score normalization to adjust each data dimension to a zero-mean, unit-variance distribution. Material balance data includes first, second, and third material balance data; equipment operating data includes electrolyzer voltage, electrolyzer current, distillation column top temperature, and distillation column bottom temperature; and environmental emission data includes hydrogen fluoride concentration in flue gas from the waste gas incineration unit and fluoride ion concentration in wastewater from the neutralization unit. For each data dimension, all observations for that dimension are extracted from the historical full-lifecycle batch data set. The arithmetic mean and standard deviation of the observations are calculated, and the original collected values ​​for the current batch are transformed using the arithmetic mean and standard deviation. The transformation formula is as follows:

[0060]

[0061] in, This represents the data value after standardization. This represents the raw data values ​​collected in the current batch. This represents the arithmetic mean of the observations in the corresponding data dimension across the entire historical batch dataset. This represents the standard deviation of the observations for the corresponding data dimension in the historical full lifecycle batch dataset. The historical full lifecycle batch dataset must include material balance data, equipment operating data, and environmental emission data for at least the 100 most recent consecutive production batches. Arithmetic mean. The calculation method is to sum all observations of the corresponding data dimension in the historical full-lifecycle batch data set and then divide by the total number of observations. Standard deviation The calculation method is as follows: first calculate the arithmetic mean of each observation. The difference is calculated by squaring the differences, summing them, dividing by the total number of observations, and taking the square root. After Z-score normalization, the numerical distribution of each data dimension satisfies the distribution characteristics of a mean of 0 and a variance of 1.

[0062] After identifying the critical control stages, process parameter optimization is performed. The process includes extracting the historical operating parameter record set corresponding to the critical control stages. The critical control stages are identified through reverse tracing and attribution. The historical operating parameter record set is extracted from the process control database according to the stage identifier and time range of the critical control stage. The historical operating parameter record set includes electrolysis temperature records, distillation reflux ratio records, and catalyst addition rate records. Each record entry in the electrolysis temperature record includes the measured temperature value and its corresponding timestamp; each record entry in the distillation reflux ratio record includes the reflux ratio setpoint and the actual measured value and their corresponding timestamps; and each record entry in the catalyst addition rate record includes the catalyst addition rate setpoint and its corresponding timestamp.

[0063] The historical operational parameter record set is associated and aligned with the feature vectors of the corresponding stages in the inter-stage material migration feature map to generate parameter-influence feature pairs. The inter-stage material migration feature map is generated by a recursive convolutional deep fusion network, with each lifecycle stage in the inter-stage material migration feature map corresponding to one dimension. The feature vectors corresponding to the key control stages are extracted from the inter-stage mass migration feature map by stage index. The association and alignment process is as follows: using the timestamp as the alignment key, the operating parameter record corresponding to each timestamp in the historical operating parameter record set is paired with the corresponding stage feature vector of the batch closest in time in the inter-stage mass migration feature map. For batch processing, the key control stage feature vector of the same batch is paired with all operating parameter records collected during the key control stages of that batch. Each pair of data includes a... The stage feature vector is used as the feature input, and a set of operating parameter values ​​is used as the output target. The operating parameter values ​​correspond to three dimensions: electrolysis temperature, distillation reflux ratio, and catalyst addition rate. The paired data of each operating parameter dimension and the stage feature vector are stored as three sets of parameter influence feature pairs.

[0064] A Bayesian optimization surrogate model is used to perform Gaussian process regression fitting on the parameter influence feature pairs. The Bayesian optimization surrogate model consists of three independent Gaussian process regression models, corresponding to the optimization objectives of electrolysis temperature, distillation reflux ratio, and catalyst addition rate, respectively. Each Gaussian process regression model includes a mean function and a covariance function. The mean function is a constant function, with its initial value set as the arithmetic mean of the target variables in the parameter influence feature pair set. The covariance function is a squared exponential kernel function, whose expression is controlled by the signal variance parameter and the length scale parameter. The signal variance parameter controls the amplitude of the function value variation, and the length scale parameter controls the rate of correlation decay between two points in the input space. The hyperparameters of the Gaussian process regression model include the constant value of the constant mean function, the signal variance parameter, and the length scale parameter corresponding to each input dimension. The hyperparameters are optimized by maximizing the logarithmic marginal likelihood function on the parameter influence feature pair set. The optimization is performed iteratively using the L-BFGS algorithm, with a maximum number of iterations set to 200.

[0065] After the Gaussian process regression model is trained, the Bayesian optimization surrogate model generates the predicted mean and variance within the parameter space to be optimized. The expected hypervolume improvement acquisition function is used to recommend the next set of process parameters to be evaluated in a multi-objective optimization scenario. The optimization objectives include two aspects: minimizing the total cumulative contribution of the critical control stage to the greenhouse effect category, and minimizing the operating energy cost of the critical control stage. The operating energy cost is calculated by comprehensively extracting electrolyzer power consumption, distillation column heating steam consumption, and catalyst procurement cost from the historical operating parameter record set. The expected hypervolume improvement acquisition function considers the improvement amounts of both optimization objectives simultaneously, calculating the hypervolume enclosed by the non-dominated solution set of the current Pareto front in the two-dimensional objective space, and evaluating the expected increase in hypervolume for each candidate process parameter combination. A closed-loop process of acquisition function recommendation and actual evaluation is iteratively executed. In each iteration, a new parameter combination and its corresponding evaluation result are added to the parameter influence feature pair set, and the Gaussian process regression model is updated, until the preset maximum number of iterations (50) is reached or the expected hypervolume improvement value is below 0.001 for five consecutive iterations.

[0066] After iteration, a Pareto front set is extracted from all evaluated process parameter combinations. Each element in the Pareto front set is a non-dominated process parameter combination. An ideal point is constructed, with coordinates in a two-dimensional target space: the first coordinate is the minimum total cumulative contribution of the greenhouse effect category among all evaluated process parameter combinations, and the second coordinate is the minimum operating energy cost among all evaluated process parameter combinations. The Euclidean distance between each process parameter combination in the Pareto front set and the ideal point is calculated, and the process parameter combination with the smallest Euclidean distance is selected as the recommended optimization parameters. The recommended optimization parameters include electrolysis temperature setpoint, distillation reflux ratio setpoint, and catalyst addition rate setpoint. The recommended optimization parameters are sent to the process control system corresponding to the key control stage in the form of process parameter adjustment commands.

[0067] In practical implementation, after outputting the process optimization anchor point, a dynamic update process for the lifecycle inventory is performed. This dynamic update process includes: acquiring measured emission data after adjusting process parameters during the critical control phase; comparing the measured emission data with the preset baseline emission data in the process optimization anchor point to generate an emission deviation vector; inputting the emission deviation vector into a preset inventory correction mapping layer, which determines the correction direction based on the sign of the emission deviation vector and the correction step size based on the absolute value of the emission deviation vector; updating the inventory factor values ​​for all stages in the entire lifecycle stage sequence, from the critical control stage to the waste disposal stage, according to the correction direction and correction step size, generating an updated lifecycle inventory database, and feeding the updated lifecycle inventory database back to the input layer of the recursive convolutional deep fusion network for inter-stage material flow tracking processing for the next batch of products.

[0068] Critical control phases are identified through reverse source attribution processing, and process optimization anchor points include the equipment identifiers and operating parameter ranges corresponding to the critical control phases. After performing process parameter adjustment operations during the critical control phases, measured emission data are collected from the emission monitoring points corresponding to the critical control phases. The measured emission data collection targets are determined based on the specific type of the critical control stage: When the critical control stage is the electrolysis for hydrofluoric acid production, the measured emission data includes the anode gas escape flow rate, the cathode liquid production flow rate, and the hydrogen fluoride concentration at the electrolytic cell exhaust outlet; when the critical control stage is the catalytic fluorination synthesis stage, the measured emission data includes the fluorinated carbon compound concentration at the reactor tail gas outlet and the mass value of spent catalyst emissions; when the critical control stage is the purification and distillation stage, the measured emission data includes the fluorine-containing gas concentration at the top of the column (non-condensable gas) and the mass value of residual liquid emissions at the bottom of the column; when the critical control stage is the ore pretreatment stage, the measured emission data includes the tailings emission mass value and dust concentration; when the critical control stage is the waste disposal stage, the measured emission data includes the hydrogen fluoride concentration at the flue gas outlet of the waste gas incineration unit, the fluoride ion concentration at the wastewater neutralization unit outlet, and the mass value of landfill material at the solid waste landfill unit. The time window for collecting measured emission data is the period after a complete batch of operations is completed in this critical control stage after the process parameters are adjusted.

[0069] The baseline emission data preset in the process optimization anchor point is extracted from the baseline emission database corresponding to the operating parameter range included in the process optimization anchor point. The baseline emission database stores the historical average emission levels under various parameter combinations within the operating parameter range during the key control stages. The historical average emission levels are calculated by weighting the emission monitoring records of at least 50 batches of normal operation according to the similarity of operating parameters. The measured emission data is compared with the baseline emission data one by one according to emission indicators to generate an emission deviation vector. The dimension of the emission deviation vector is equal to the number of emission indicators in the measured emission data. The calculation method of each component in the emission deviation vector is: the measured value of the k-th emission indicator minus the baseline value of the k-th emission indicator, and then divided by the baseline value of the k-th emission indicator to obtain the relative deviation of the k-th emission indicator. Each component of the emission deviation vector is a dimensionless value. The sign of the component indicates the direction of increase or decrease of the emission relative to the baseline value. A positive value indicates that the emission is higher than the baseline value, and a negative value indicates that the emission is lower than the baseline value. The absolute value of the component indicates the magnitude of the deviation of the emission relative to the baseline value.

[0070] The emission deviation vector is input into a predefined inventory correction mapping layer. The inventory correction mapping layer receives the emission deviation vector as input and outputs an inventory factor correction coefficient vector. Each component in the inventory factor correction coefficient vector corresponds to the inventory factor correction coefficient for one stage in the life cycle stage sequence. Inventory factors include raw material consumption factors, energy consumption factors, greenhouse gas emission factors, acid gas emission factors, ecotoxicity emission factors, and waste generation factors. The inventory correction mapping layer is a fixed mapping matrix, with the number of rows equal to the total number of stages in the life cycle stage sequence, and the number of columns equal to the dimension of the emission deviation vector. The element value in the p-th row and k-th column of the mapping matrix represents the weight of the deviation of the k-th emission indicator on the inventory factor correction coefficient for the p-th life cycle stage. The influence weights are determined as follows: During the process design phase, a steady-state simulation of the perfluorotributylamine production process is performed. Unit step changes in emission indicators at each key control stage are introduced as disturbances. These disturbances are propagated in the steady-state simulation model, and the relative changes in inventory factor values ​​for all stages of the entire lifecycle are recorded relative to the unperturbed state. These relative changes are normalized and used as the influence weight values ​​for the corresponding columns of the mapping matrix. The sum of the influence weight values ​​in each column of the mapping matrix is ​​1.

[0071] The inventory correction mapping layer determines the correction direction based on the sign of the emission deviation vector by performing matrix multiplication between the emission deviation vector and the mapping matrix to generate an intermediate correction vector. The dimension of the intermediate correction vector is the total number of stages in the entire life cycle sequence. The sign of the p-th component in the intermediate correction vector determines the correction direction of the inventory factor for the p-th life cycle stage; a positive sign indicates that the inventory factor value should be increased, and a negative sign indicates that the inventory factor value should be decreased.

[0072] The inventory correction mapping layer determines the correction step size based on the absolute value of the emission deviation vector as follows: The absolute value of each component in the emission deviation vector is multiplied by the influence weight of the corresponding column in the mapping matrix, and then summed to obtain a correction magnitude scalar for each life cycle stage. This correction magnitude scalar is then multiplied by a set maximum correction step size constant to obtain the correction step size for each life cycle stage. The maximum correction step size constant is set to 0.05. The rationale for setting this constant is that in life cycle assessment data quality grading, the uncertainty of inventory data is typically controlled within the range of 5% to 10%. Selecting 5% as the maximum step size for a single adjustment strikes a balance between data update stability and convergence speed.

[0073] Once the correction direction and step size are determined, update the inventory factor values ​​for all stages in the lifecycle stage sequence, from the critical control stage to the waste disposal stage, according to the correction direction and step size. The update operation uses the following formula:

[0074]

[0075] in, This represents the updated list factor value for the p-th lifecycle stage. The value range is from the stage number of the critical control stage to the stage number of the waste disposal stage; This represents the list factor value of the p-th lifecycle stage before the update, which is read from the lifecycle list database before this update. This represents the correction step size for the p-th lifecycle stage. The value of is determined by the correction step size output by the list correction mapping layer, under the constraint of a maximum correction step size constant of 0.05. The value of is a real number between 0 and 0.05; This represents the correction direction indicator for the p-th lifecycle stage. The value can be +1 or -1, where +1 indicates an upward adjustment of the list factor and -1 indicates a downward adjustment of the list factor. The value of is determined by the sign direction of the p-th component of the intermediate correction vector.

[0076] After updating all inventory factor values ​​for all stages from the critical control stage to the waste disposal stage according to the above formula, the updated inventory factor values ​​are written back to the lifecycle inventory database to generate the updated lifecycle inventory database.

[0077] The updated lifecycle inventory database is fed back to the input layer of the recursive convolutional deep fusion network for inter-stage material flow tracking processing of the next batch of products. The feedback method is as follows: when the full lifecycle management process for the next batch of products is initiated, the input layer of the recursive convolutional deep fusion network, while collecting material balance data, equipment operating data, and environmental emission data, reads the updated inventory factor values ​​from the updated lifecycle inventory database and appends these inventory factor values ​​as additional channels to the multi-channel material flow tensor. The original three channels of the multi-channel material flow tensor are the first material balance data, the second material balance data, and the third material balance data. The additional channels are a one-dimensional sequence of the updated inventory factor value vector for each lifecycle stage after dimensionality reduction. The addition of additional channels does not change the internal structure of the spatial convolutional layer and the long short-term memory layer of the recursive convolutional deep fusion network; the number of channels in the convolutional kernel of the spatial convolutional layer is increased accordingly to match the total number of channels in the multi-channel material flow tensor. After receiving input from the additional inventory factor channel, the recursive convolutional deep fusion network extracts local migration feature maps and inter-stage material migration feature maps, which implicitly contain information about the previous batch of lifecycle inventory updates, thus achieving a data closed loop for the lifecycle management of subsequent batches of products.

[0078] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A product lifecycle management method based on perfluorotributylamine, characterized in that, The method includes: Obtain the full life cycle sequence of perfluorotributylamine products from raw material input to final disposal. The full life cycle sequence includes the ore pretreatment stage, the electrolytic hydrofluoric acid production stage, the catalytic fluorination synthesis stage, the purification and distillation stage, the product application stage, and the waste disposal stage. For each life cycle stage, material balance data, equipment operating condition data and environmental emission data are collected, and the material balance data is input into a recursive convolutional deep fusion network to perform cross-stage material flow tracking processing to generate inter-stage material migration feature maps. The pre-trained life cycle impact assessment model is invoked to perform multi-objective environmental impact decoupling processing on the inter-stage material migration feature map, and output the environmental impact contribution vector corresponding to each life cycle stage. Based on the environmental impact contribution vector, reverse attribution processing is performed on the full life cycle stage sequence to mark at least one key control stage whose contribution exceeds the dynamic threshold, and the key control stage is output as the process optimization anchor point.

2. The product lifecycle management method based on perfluorotributylamine according to claim 1, characterized in that, The complete lifecycle sequence of obtaining perfluorotributylamine products, from raw material input to final disposal, includes: Collect the mining batch identifier and transportation route information of fluorite raw materials, and allocate the fluorite raw materials to the input queue of the ore pretreatment stage according to the mining batch identifier; The start-up and shutdown events of the electrolyzer in the electrolytic hydrofluoric acid production stage are modeled using a state machine to generate an electrolysis process state transition sequence, and the electrolysis process state transition sequence is aligned and bound with the feed sequence of the catalytic fluorination synthesis stage. The top and bottom streams of the multi-stage distillation column in the purification and distillation stage are subjected to component identification and tracking processing to generate component flow direction tags, and the component flow direction tags are attached to the input material descriptor in the product application stage. The operation logs of the waste gas incineration device, wastewater neutralization device and solid waste landfill device in the waste disposal stage are timestamped and aligned. The aligned operation logs are then linked with the output residual streams in the product application stage.

3. The product lifecycle management method based on perfluorotributylamine according to claim 1, characterized in that, The process involves collecting material balance data, equipment operating condition data, and environmental emission data for each lifecycle stage, and inputting the material balance data into a recursive convolutional deep fusion network to perform cross-stage material flow tracing processing, generating inter-stage material migration feature maps, including: The difference in the inflow and outflow quality of fluorite tailings during the ore pretreatment stage is collected as the first material balance data. The ratio of anode gas emission to cathode liquid output during the electrolytic hydrofluoric acid production stage is collected as the second material balance data. The ratio of perfluorotributylamine crude product generation to catalyst consumption during the catalytic fluorination synthesis stage is collected as the third material balance data. The first material balance data, the second material balance data, and the third material balance data are spliced ​​together in the stage time sequence to form a multi-channel material flow tensor. The multi-channel material flow tensor is input into the spatial convolutional layer of the recursive convolutional deep fusion network to perform local material migration pattern extraction processing, generating a local migration feature map. The local migration feature map is input into the long short-term memory layer of the recursive convolutional deep fusion network for cross-stage dependency encoding to generate the inter-stage material migration feature map.

4. The product lifecycle management method based on perfluorotributylamine according to claim 3, characterized in that, The local migration feature map is input into the long short-term memory layer of the recursive convolutional deep fusion network for cross-stage dependency encoding to generate the inter-stage material migration feature map, including: The local migration feature map corresponding to the i-th life cycle stage is input into the forgetting gate unit of the long short-term memory layer. The forgetting gate unit reads the hidden state vector of the previous stage and the input feature vector of the current stage, and outputs the material discarding weight matrix between stages. The local migration feature map is input into the input gating unit of the long short-term memory layer. The input gating unit calculates the candidate memory cell vector and updates the memory cell state of the current stage according to the inter-stage material discarding weight matrix. The updated memory cell state is input into the output gating unit of the Long Short-Term Memory layer, and the output gating unit generates the hidden state vector of the current stage as the feature representation of the corresponding stage in the inter-stage material migration feature map.

5. The product lifecycle management method based on perfluorotributylamine according to claim 1, characterized in that, The pre-trained lifecycle impact assessment model is invoked to perform multi-objective environmental impact decoupling processing on the inter-stage material migration feature map, outputting an environmental impact contribution vector corresponding to each lifecycle stage, including: The inter-stage material migration feature map is input into the global average pooling layer of the life cycle impact assessment model for feature dimensionality reduction to generate stage aggregated feature vectors. The aggregated feature vectors of the aforementioned stages are input into the multi-head attention decomposition layer of the life cycle impact assessment model. The multi-head attention decomposition layer includes a greenhouse effect attention head, an acidification effect attention head, and an ecotoxicity attention head. Each attention head independently calculates the impact weight distribution of each life cycle stage. The environmental impact contribution vector is generated by concatenating the first impact weight vector output by the greenhouse effect attention head, the second impact weight vector output by the acidification effect attention head, and the third impact weight vector output by the ecotoxicity attention head.

6. The product lifecycle management method based on perfluorotributylamine according to claim 5, characterized in that, The aggregated feature vectors of the aforementioned stages are input into the multi-head attention decomposition layer of the life cycle impact assessment model. This multi-head attention decomposition layer includes a greenhouse effect attention head, an acidification effect attention head, and an ecotoxicity attention head. Each attention head independently calculates the impact weight distribution for each life cycle stage, including: The stage aggregated feature vector is multiplied by the first query weight matrix, the first key weight matrix and the first value weight matrix of the greenhouse effect attention head to generate a greenhouse effect query vector, a greenhouse effect key vector and a greenhouse effect value vector. Calculate the dot product similarity matrix between the greenhouse effect query vector and the greenhouse effect key vector, input the dot product similarity matrix into the flexible maximum value normalization unit of the greenhouse effect attention head for inter-stage weight normalization processing, and generate the first influence weight vector; The greenhouse effect value vector is weighted and summed with the first impact weight vector to generate a greenhouse effect weighted feature vector, and the greenhouse effect weighted feature vector is passed to the output splicing layer of the life cycle impact assessment model.

7. The product lifecycle management method based on perfluorotributylamine according to claim 1, characterized in that, Based on the environmental impact contribution vector, reverse attribution processing is performed on the entire life cycle stage sequence to identify at least one key control stage whose contribution exceeds a dynamic threshold. This key control stage is then output as an anchor point for process optimization, including: For each component in the environmental impact contribution vector, a reverse traversal scan is performed from the product application stage to the ore pretreatment stage, and the environmental impact contribution of the current stage is distributed in reverse to the input material node of the previous stage according to the material migration path. The dynamic threshold for each life cycle stage is calculated based on the cumulative contribution value after reverse allocation, and the dynamic threshold is twice the average contribution value of the entire life cycle stage sequence. The lifecycle stages whose contribution exceeds the dynamic threshold are marked as the key control stages, and the device identifiers and operating parameter ranges corresponding to the key control stages are extracted as the process optimization anchors.

8. The product lifecycle management method based on perfluorotributylamine according to claim 7, characterized in that, Perform a reverse traversal scan from the product application stage to the ore pretreatment stage for each component of the environmental impact contribution vector, and distribute the environmental impact contribution of the current stage in reverse order to the input material nodes of the previous stage according to the material migration path, including: Obtain the material migration coefficient matrix between stage j and stage j minus one in the full life cycle stage sequence. The row index of the material migration coefficient matrix corresponds to the output material type of stage j, and the column index corresponds to the input material type of stage j minus one. Perform an inverse matrix multiplication operation between the environmental impact contribution vector of stage j and the material migration coefficient matrix to generate the contribution vector to be allocated for stage j minus one. The contribution vector to be assigned is added to the original direct environmental impact contribution vector of the j-th minus one stage to update the cumulative contribution vector of the j-th minus one stage. Repeat the above steps until the ore pretreatment stage is reached, generating the final reverse cumulative contribution for each lifecycle stage.

9. The product lifecycle management method based on perfluorotributylamine according to claim 1, characterized in that, The material balance data, equipment operating data, and environmental emission data are standardized before being input into the recursive convolutional deep fusion network. The standardization process uses the Z-score normalization method to adjust each data dimension to a zero-mean, unit variance distribution.

10. A perfluorotributylamine product lifecycle management system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the perfluorotributylamine product lifecycle management method as described in any one of claims 1 to 9.

Images