An ai-enabled supercritical foam production anomaly monitoring management system

Abstract

The application discloses an AI-enabled supercritical foaming material production anomaly monitoring and management system, and belongs to the technical field of production management, and specifically comprises: a log input module, which inputs a non-parametric production event log file; a vector extraction module, which obtains a customer code list and extracts a complaint event frequency distribution vector; a segment coding module, which divides production events into multiple production segment intervals and extracts an event type coding sequence; a risk mapping module, which inputs the coding sequence into a complaint risk mapping model to output complaint risk tendency coding; a matching and distribution module, which inputs the complaint vector and the risk coding into a distribution rule library to output a distribution binding relationship record; and an outbound identification module, which stores a finished product packaging box identification code in association with a customer code and adds a risk code in a delivery document. The application realizes the mapping between production event sequences and customer complaint risks and the differentiated distribution of finished products.

Description

Technical Field

[0001] This invention relates to the field of production management technology, specifically to an AI-enabled monitoring and management system for abnormal production of supercritical foamed materials. Background Technology

[0002] Supercritical foam materials are a class of high-performance, lightweight materials that use supercritical fluids as physical foaming agents to form microporous structures in polymer melts. They are widely used in sports shoe midsoles, automotive interiors, and electronic cushioning packaging. The production process of supercritical foam materials is continuous and long-cycle, resulting in cumulative microscopic differences in foaming ratio uniformity and surface quality within the same production batch over time. In supply chain management practice, finished products from the same production batch are often stored together and shipped to multiple downstream customers. Each customer's tolerance for product quality fluctuations varies significantly depending on their application scenarios and end-market positioning.

[0003] Existing production anomaly monitoring and management systems primarily rely on parameterized data streams such as pressure and temperature sensor readings to monitor the production process of supercritical foamed materials. Anomaly alarms and batch quality assessments are triggered by setting upper and lower thresholds for these parameters. In the finished product allocation stage, existing enterprise resource planning (ERP) systems typically allocate finished products based on order delivery dates or customer geographic distribution, without considering the matching relationship between product quality differences across different production time segments within a batch and customer quality tolerance. Some management systems introduce statistical process control methods to evaluate overall batch quality indicators, but the evaluation results are applied uniformly to the entire batch.

[0004] However, in the production of supercritical foamed materials, human intervention by operators, fluctuations in operating conditions caused by shift changes, batch differences in raw materials due to material feeding changes, and residual system state deviations after equipment alarm resets are all recorded as non-parametric events in the event log of the production execution system. The combination and arrangement patterns of these discrete events on the timeline imply the cumulative trajectory of process state deviations, and the risk of complaints from customers in the subsequent use of finished products corresponding to different event combinations varies quantifiablely. However, existing systems only use event logs as auxiliary records for production traceability and do not incorporate them into the quality prediction and order allocation decision-making process. Parametric threshold alarm mechanisms can only identify real-time out-of-limit conditions and cannot analyze the risk tendencies implied in the event sequence; the unified allocation strategy for the entire batch ignores the matching space between the risk differences between different production segments within the same batch and the different customer complaint tolerance levels. This technological gap means that the objectively existing event combination risks in the production process cannot be transmitted to the finished product outbound allocation stage. High-risk segment finished products may flow to customers with lower complaint tolerance, while low-risk segment finished products fail to prioritize the needs of highly sensitive customers, resulting in a structural waste of supply chain quality costs. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-enabled monitoring and management system for anomalies in the production of supercritical foamed materials, addressing the following technical problems:

[0006] Existing technologies fail to establish a mapping relationship between non-parametric production event sequences and customer complaint risks, and cannot distinguish the risk tendencies of finished products by production segments and match the complaint tolerance of different customers, resulting in a disconnect between finished product outbound allocation decisions and quality risks.

[0007] The objective of this invention can be achieved through the following technical solutions: An AI-powered monitoring and management system for anomalies in the production of supercritical foamed materials includes: The log input module is used to input a non-parametric production event log file for the production batch of supercritical foamed materials. The non-parametric production event log file contains production event entries recorded in timestamp order. The vector extraction module is used to retrieve the product sales flow database based on the batch number of the production batch, obtain the customer code list, retrieve the customer historical complaint record database for each customer code, and extract the frequency distribution vector of complaint events. The segment encoding module is used to divide production event entries into multiple production segment intervals according to the production scheduling timeline, and extract the event type encoding sequence within each production segment interval. The risk mapping module is used to input the event type coding sequence of each production segment interval into the production segment complaint risk mapping model. The production segment complaint risk mapping model is constructed based on the co-occurrence relationship between historical event coding combinations and complaint records, and outputs a complaint risk tendency code. The matching and allocation module is used to input the frequency distribution vector of complaint events for each customer code and the complaint risk tendency code for each production segment into the customer order allocation rule base, and output the allocation binding relationship record after performing the matching operation. The outbound identification module is used to associate the finished product packaging box identification code with the customer code in the outbound shipment details table during the finished product outbound stage of the enterprise resource planning system, based on the allocation binding relationship record, and to attach a complaint risk tendency code to the shipment document.

[0008] As a further aspect of the present invention: the specific process of inputting the non-parametric production event log file of the supercritical foaming material production batch in the log input module is as follows: The non-parametric production event log file is generated by the production execution system one by one according to the event trigger time during the operation of the supercritical foaming material production line. The production event entries include equipment start-up and shutdown event entries, manual intervention operation event entries, production shift handover event entries, material feeding switching event entries, and equipment alarm reset event entries. Each production event entry also includes an event source equipment code field and an event supplementary description text field. The event supplementary description text field is written by obtaining the event type code fields of all production event entries within a preset time window before the event occurrence timestamp and concatenating them in reverse chronological order to form an event prefix code string. Then, the event type code fields of all production event entries within a preset time window after the event occurrence timestamp are obtained and concatenated in ascending chronological order to form an event suffix code string. The event prefix code string and the event suffix code string are connected with a preset separator and then written into the event supplementary description text field.

[0009] As a further aspect of the present invention: the specific process of retrieving the customer's historical complaint record database and extracting the frequency distribution vector of complaint events for each customer code in the vector extraction module is as follows: Retrieve all complaint records generated by the customer code within the historical time period from the customer history complaint record database. Each complaint record contains a complaint issue type code field and a complaint occurrence date field. After deduplication of all complaint issue type codes that occurred within the historical time period, sort them in ascending order by code value to form a complaint issue type code dimension list. For each complaint issue type code in the complaint issue type code dimension list, obtain the set of complaint records corresponding to the complaint issue type code of the customer code within the historical time period. Assign a time decay weight value to each complaint record in the complaint record set. The time decay weight value is calculated by calculating the date difference between the complaint occurrence date field and the base date using the current system date as the base date. Input the date difference into a preset negative exponential decay function to obtain the time decay weight value of the complaint record. Accumulate the time decay weight values ​​of all complaint records in the complaint record set, and determine the accumulated result as the vector component value corresponding to the dimension in the complaint event frequency distribution vector.

[0010] As a further aspect of the present invention: the specific process of dividing production event entries into multiple production segment intervals according to the production scheduling timeline and extracting the event type encoding sequence within each production segment interval in the segment encoding module is as follows: Obtain the event occurrence timestamps of all production shift handover event entries recorded on the production scheduling timeline for the production batch. Determine the time span between the event occurrence timestamps of two adjacent production shift handover event entries as the duration of a production segment interval. Each production segment interval corresponds to a segment number. Extract the event type code field and the event source equipment code field of all production event entries occurring within each production segment interval. Arrange the event type code field and the event source equipment code field in an alternating manner according to the chronological order of the event occurrence timestamps of their respective corresponding production event entries to form an event code hybrid sequence. Determine the event code hybrid sequence as the event type code sequence corresponding to the production segment interval.

[0011] As a further aspect of the present invention: the specific process of inputting the event type encoding sequence of each production segment interval into the production segment complaint risk mapping model and outputting the complaint risk tendency code in the risk mapping module is as follows: The event-coded mixed sequence of each historical production segment interval in the historical production batch set is obtained and combined with the customer complaint result identifier field to form a training sample data record. The weighted edit distance value between every two event-coded mixed sequences in the training sample data record set is calculated. A hierarchical clustering tree structure is constructed based on the weighted edit distance value. The proportion value of the complaint result identifier field is counted at each cluster node of the hierarchical clustering tree structure. When the proportion value exceeds a preset proportion threshold, the event-coded mixed subsequence with the highest frequency in the cluster node is marked as a complaint-related feature subsequence and assigned a corresponding complaint risk tendency code. The event-coded mixed sequence of the current production segment interval is matched with each complaint-related feature subsequence, and the complaint risk tendency code corresponding to the complaint-related feature subsequence with the highest matching degree is determined as the output.

[0012] As a further aspect of the present invention: the weighted edit distance value is calculated as follows: the replacement cost between event type codes is determined based on the event classification hierarchy relationship stored in the production event classification hierarchy table; the replacement cost between event source equipment codes is determined based on the upstream and downstream location distance of the equipment stored in the production line equipment topology relationship table; when calculating the proportion of the complaint result identifier field, the training sample data records in the cluster node are first grouped by customer code, the ratio of the number of complaint records in each customer code group to the total number of training sample data records is calculated, and then the arithmetic mean of the ratios of each customer code group is calculated.

[0013] As a further aspect of the present invention: the specific process of outputting the allocation binding relationship record after performing the matching operation in the matching and allocation module is as follows: Customer codes are clustered and grouped based on the cosine distance of the frequency distribution vector of complaint events for each customer code within a historical time period, and a risk tolerance threshold range is set for each customer group. After each production batch is delivered, the complaint risk tendency code for the production segment range assigned to the customer code is obtained. If the complaint risk tendency code exceeds the risk tolerance threshold range and the customer code has not complained, a positive adjustment cumulative value is recorded. If the complaint risk tendency code falls within the risk tolerance threshold range and the customer code complains, a negative adjustment cumulative value is recorded. When the positive adjustment cumulative value exceeds the preset positive adjustment threshold, the customer group to which the customer code belongs is re-determined. When the negative adjustment cumulative value exceeds the preset negative adjustment threshold, the upper and lower boundary values ​​of the risk tolerance threshold range of the customer group are narrowed. When performing the matching operation, the complaint risk tendency code of the current production segment range is compared with the risk tolerance threshold range of the customer group to which the customer code to be assigned belongs. If they fall within the range, the allocation binding relationship record is output.

[0014] As a further aspect of the present invention: the specific process of adding a complaint risk propensity code to the delivery document in the outbound identification module is as follows: The warehouse execution system prints a two-dimensional barcode graphic containing the customer code and complaint risk propensity code on the logistics label based on the correspondence between the finished product packaging box identification code and the customer code in the outbound shipment details table. When the handheld scanning terminal reads the two-dimensional barcode graphic, it sends a query request containing the complaint risk propensity code and the current location information to the enterprise resource planning system. The enterprise resource planning system retrieves the historical complaint problem type distribution frequency based on the complaint risk propensity code, and retrieves the historical cargo damage record number through transit warehouses based on the current location information. It then inputs the historical complaint problem type distribution frequency and the historical cargo damage record number into the acceptance prompt text generation template to generate an unloading acceptance prompt text, which is then returned to the handheld scanning terminal for display.

[0015] The beneficial effects of this invention are: This invention inputs a non-parametric production event log file and extracts the event type coding sequence. Production batches are divided into multiple production segment intervals based on the handover timestamps of production shifts. A weighted edit distance algorithm is used to perform hierarchical clustering of the mixed event coding sequences of historical production segments and the customer complaint result identifier field, constructing a production segment complaint risk mapping model and outputting the complaint risk tendency code for each production segment interval. Simultaneously, a customer complaint event frequency distribution vector is extracted based on a time decay weighting method. Customers are clustered and grouped, and a dynamically adjusted risk tolerance threshold interval is set. Through matching operations, production segments whose complaint risk tendency codes fall within the customer risk tolerance threshold interval are assigned and bound to corresponding customers. During the finished product outbound stage in the enterprise resource planning system, the finished product packaging box identifier code is associated with the customer code and stored, and the complaint risk tendency code is attached to the logistics label for differentiated acceptance in the logistics process. This establishes a complete data flow from production event sequences to customer complaint risk mapping, from risk mapping to customer tolerance matching, and from matching results to finished product graded outbound, solving the deficiency in existing technologies where the implicit risks of production events cannot be quantified and transmitted to the order allocation decision-making stage. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a schematic diagram of the modules of the present invention. Detailed Implementation

[0018] 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.

[0019] Please see Figure 1As shown, this invention is an AI-enabled monitoring and management system for anomalies in the production of supercritical foamed materials, comprising: The log input module is used to input a non-parametric production event log file for the production batch of supercritical foamed materials. The non-parametric production event log file contains production event entries recorded in timestamp order. The vector extraction module is used to retrieve the product sales flow database based on the batch number of the production batch, obtain the customer code list, retrieve the customer historical complaint record database for each customer code, and extract the frequency distribution vector of complaint events. The segment encoding module is used to divide production event entries into multiple production segment intervals according to the production scheduling timeline, and extract the event type encoding sequence within each production segment interval. The risk mapping module is used to input the event type coding sequence of each production segment interval into the production segment complaint risk mapping model. The production segment complaint risk mapping model is constructed based on the co-occurrence relationship between historical event coding combinations and complaint records, and outputs a complaint risk tendency code. The matching and allocation module is used to input the frequency distribution vector of complaint events for each customer code and the complaint risk tendency code for each production segment into the customer order allocation rule base, and output the allocation binding relationship record after performing the matching operation. The outbound identification module is used to associate the finished product packaging box identification code with the customer code in the outbound shipment details table during the finished product outbound stage of the enterprise resource planning system, based on the allocation binding relationship record, and to attach a complaint risk tendency code to the shipment document.

[0020] In a preferred embodiment of the present invention, the specific process of inputting the non-parametric production event log file of the supercritical foaming material production batch in the log input module is as follows: The log input module receives non-parametric production event log files generated sequentially by the production execution system during the operation of the supercritical foaming material production line, according to the event trigger time. Each production event entry in the log file includes an event occurrence timestamp field, an event type code field, an event source equipment code field, and an event supplementary description text field. The event occurrence timestamp field records the precise time of event triggering, with a time precision of milliseconds, for example, 14:23:45.678 on November 17, 2025. The event type code field is assigned values ​​using a preset coding table, which divides event types into five categories: equipment start / stop events (code range E001 to E099), manual intervention operation events (code range H001 to H099), production shift handover events (code range S001 to S099), material feeding switching events (code range M001 to M099), and equipment alarm reset events (code range A001 to A099). Specifically, the equipment start-up event code is E001, and the equipment stop-up event code is E002; the event code for the operator modifying the supercritical fluid injection pressure setting value through the human-machine interface is H011, and the event code for modifying the constant temperature holding time setting value of the foaming kettle is H021; the handover event code between the morning shift and the afternoon shift is S001, and the handover event code between the afternoon shift and the evening shift is S002; the event code for switching the supercritical fluid raw material tank from batch number C20251011 to batch number C20251025 is M003; and the event code for the operator performing a reset operation after the overpressure alarm signal of the foaming kettle is generated is A004. The Event Source Equipment Code field records the unique identifier code of the physical equipment that triggered the event in the equipment management ledger. For example, the equipment code for the supercritical fluid injection unit is SCF-INJ-03, the equipment code for the foaming vessel unit is FOAM-VSL-07, the equipment code for the pressure relief control unit is DEPR-CTL-12, and the equipment code for the melt pump unit is MELT-PMP-05.

[0021] The writing process for the event supplementary description text field adopts an encoding string concatenation method based on the event context window. For any production event entry, the event occurrence timestamp of the event entry is first obtained and recorded as the base timestamp. Then, a preset time window length is traced backwards in the negative direction of the time axis. The preset time window length is set to 180 seconds, and all production event entries that occurred within this time window are extracted. These production event entries are arranged in reverse order from latest to earliest according to their event occurrence timestamps. The event type encoding field of each production event entry is extracted sequentially. The extracted event type encoding fields are concatenated end to end using a hyphen as a connector. The concatenation result is recorded as the event prefix encoding string. For example, if three events, E002, H011, and S001, occurred sequentially within 180 seconds before the base timestamp, the event prefix encoding string would be S001-H011-E002. Next, look forward along the timeline for a preset time window length, also set to 180 seconds. Extract all production event entries occurring within this time window. Arrange these production event entries in ascending order of their event timestamps, and extract the event type code field for each production event entry. Concatenate the first and last events using a hyphen as the connector, and record the concatenated result as the event suffix code string. For example, if events M003, A004, and E001 occur sequentially within 180 seconds after the base timestamp, the event suffix code string would be M003-A004-E001. Finally, concatenate the event prefix code string and the event suffix code string using a hash symbol as the preset separator to form a complete string, which is then written into the event supplementary description text field. For example, in the above example, the final content written would be S001-H011-E002#M003-A004-E001. This writing method ensures that the supplementary description text of each production event entry carries the event sequence context information of the neighborhood before and after its occurrence, providing a structured coding basis for subsequent segmentation and sequence alignment.

[0022] In another preferred embodiment of the present invention, the specific process of retrieving the customer's historical complaint record database and extracting the frequency distribution vector of complaint events for each customer code in the vector extraction module is as follows: The vector extraction module retrieves and processes complaint records stored in the customer history complaint database to extract the frequency distribution vector of complaint events corresponding to each customer code. Each complaint record in the customer history complaint database contains three core fields: the customer code field, such as CUST-08-2217; the complaint issue type code field, which adopts the company's unified quality complaint classification coding system, such as surface defects coded as Q001, dimensional deviations coded as Q002, mechanical performance non-compliance coded as Q003, uneven foaming ratio coded as Q004, and packaging damage coded as Q005; and the complaint occurrence date field, which records the calendar date on which the customer officially submitted the complaint, such as August 14, 2025.

[0023] The vector extraction module first uses the current system date as the base date to retrieve all complaint records generated by the customer code within a historical time period, which is set to 730 days prior to the base date. The retrieved results constitute the complaint record set for that customer code. The module then iterates through the complaint issue type code field of all complaint records in this set, extracts the complaint issue type codes that have appeared, performs deduplication, and then sorts them in ascending order of code values ​​to form a complaint issue type code dimension list. Assuming that a customer code's complaint records involve three issue types: Q001, Q003, and Q004, its dimension list would be [Q001, Q003, Q004].

[0024] For each complaint type code in the dimension list, the module filters all complaint records matching that code from the complaint record set, forming a subset of complaint records corresponding to that dimension. The module calculates a time decay weight value for each complaint record in this subset. The time decay weight value is calculated by subtracting the date value in the complaint occurrence date field from the current system date, resulting in a date difference in days. This date difference is then input into a preset negative exponential decay function, where the base is the natural constant e, and the exponent is the date difference divided by the negative of the decay half-life constant. The decay half-life constant is set to 180 days. For example, if a complaint record occurs on March 1, 2025, and the current system date is November 17, 2025, the date difference is 261 days, then the time decay weight value is equal to e^(-261) divided by 180. Another complaint record occurs on October 1, 2025, with a date difference of 47 days, and its time decay weight value is significantly greater than the aforementioned record. This calculation method results in a lower contribution of older complaint records to the current vector, and a higher contribution of more recent complaint records to the current vector.

[0025] The module sums the time decay weights of all complaint records in the complaint record subset. The sum is used as the vector component value corresponding to the complaint issue type encoding dimension in the complaint event frequency distribution vector. After traversing all dimensions in the dimension list, the module obtains a numerical sequence with the same length as the dimension list. This numerical sequence is the complaint event frequency distribution vector for that customer code. For example, the vector component value of customer code CUST-08-2217 is 3.27 in dimension Q001, 1.18 in dimension Q003, and 0.93 in dimension Q004. The vector component values ​​for other dimensions that do not appear are filled with zero values.

[0026] In another preferred embodiment of the present invention, the specific process of dividing production event entries into multiple production segment intervals according to the production scheduling timeline and extracting the event type encoding sequence within each production segment interval in the segment encoding module is as follows: First, the module retrieves all production shift handover event entries recorded on the production scheduling timeline for the target production batch from the production execution system. The event type code field for each production shift handover event entry ranges from S001 to S099; for example, the handover code for the morning shift and afternoon shift is S001, and the handover code for the afternoon shift and evening shift is S002. The module then extracts the event occurrence timestamp field from all production shift handover event entries and arranges these timestamps in chronological order into a timestamp sequence. The first element in this timestamp sequence is the start timer for the batch production handover event, and the last element is the end timer for the batch production handover event. The time span between two adjacent timestamps is defined as the duration of a production segment interval.

[0027] Taking a batch of supercritical foamed material production as an example, the sequence of timestamps for the handover events between production shifts on its production scheduling timeline is as follows: first handover point 08:00:00, second handover point 16:00:00, third handover point 00:00:00, and fourth handover point 08:00:00. This batch is then divided into three production segments: the first segment corresponds to 08:00:00 to 16:00:00, lasting 8 hours; the second segment corresponds to 16:00:00 to 00:00:00 the next day, lasting 8 hours; and the third segment corresponds to 00:00:00 to 08:00:00 the next day, lasting 8 hours. Each production segment is assigned a unique segment number, which is a combination of the batch number and a two-digit serial number, such as BATCH1124-01, BATCH1124-02, and BATCH1124-03.

[0028] For each production segment interval, the module retrieves all production event entries occurring within that time span. The search scope includes the start timestamp of the segment interval but excludes the end timestamp. The module extracts the event type code field and the event source device code field for each production event entry from the search results. The event type code field is designated as the first type of code, and the event source device code field as the second type of code. The module then interleaves these two types of codes according to the chronological order of the event occurrence timestamps of their respective production event entries. The arrangement rule is as follows: for production event entries with the same timestamp, the first type of code is arranged first, followed by the second type of code; for production event entries with different timestamps, the two types of codes for the corresponding events are retrieved strictly in ascending order of timestamps. After the interleave, a linear sequence consisting of alternating first and second type codes is obtained; this linear sequence is the event code hybrid sequence.

[0029] Taking the first segment, BATCH1124-01, as an example, the following events occurred in chronological order within this segment: 08:05:22, the supercritical fluid injection unit started (equipment code SCF-INJ-03); 08:30:17, the operator modified the injection pressure setting (station code OP-STA-02); 10:45:09, the overpressure alarm of the foaming vessel unit was reset (equipment code FOAM-VSL-07); 13:20:33, the material tank was switched (tank number TANK-06). The first-class codes for each event are E001, H011, A004, and M003, respectively, and the second-class codes are SCF-INJ-03, OP-STA-02, FOAM-VSL-07, and TANK-06, respectively. The event-coded mixed sequence obtained by arranging the sequences in chronological order is as follows: E001, SCF-INJ-03, H011, OP-STA-02, A004, FOAM-VSL-07, M003, TANK-06. This event-coded mixed sequence is the event type coding sequence corresponding to this production segment interval and will be used as input feature data for the subsequent production segment complaint risk mapping model. For each production segment interval, the module generates an event-coded mixed sequence in the above manner and stores the segment number in a one-to-one correspondence with the sequence.

[0030] In another preferred embodiment of the present invention, the specific process of inputting the event type encoding sequence of each production segment interval into the production segment complaint risk mapping model and outputting the complaint risk tendency code in the risk mapping module is as follows: The risk mapping module receives the event code mixed sequence of each production segment interval output by the segment encoding module, and calls the pre-built production segment complaint risk mapping model for processing, outputting the complaint risk tendency code corresponding to each production segment interval. The construction process of the production segment complaint risk mapping model is based on the set of historical production batches of supercritical foamed materials delivered within a historical time period.

[0031] First, data is extracted batch by batch from the historical production batch set. For each historical production batch, the complaint result identifier field of each receiving customer after delivery is obtained. The complaint result identifier field is a binary variable, with a value of 1 indicating that the customer submitted a complaint about the finished product of this batch, and a value of 0 indicating that the customer did not submit a complaint. At the same time, the non-parametric production event log file generated during the production process of this historical production batch is obtained. According to the same division method as the fragment coding module, all production event entries of this batch are divided into multiple historical production fragment intervals, and a corresponding event coding mixed sequence is generated for each historical production fragment interval. The event coding mixed sequence of each historical production fragment interval and the corresponding customer complaint result identifier field are combined to form a training sample data record. For example, the event-coded mixed sequence corresponding to the third segment interval BATCH0822-03 of historical batch BATCH0822 is E001, SCF-INJ-03, H011, OP-STA-02, A004, FOAM-VSL-07, M003, TANK-06. After this batch was sent to customer code CUST-11-0532, the customer submitted a complaint. Therefore, the content of this training sample data record is: the event-coded mixed sequence E001, SCF-INJ-03, H011, OP-STA-02, A004, FOAM-VSL-07, M003, TANK-06 and the complaint result identifier field 1. All training sample data records corresponding to all historical production segment intervals of all historical production batches are collected to form a training sample data record set.

[0032] After the training sample data record set is constructed, the risk mapping module calculates the weighted edit distance between every two event-coded hybrid sequences in the set. The weighted edit distance is defined as the minimum cumulative edit cost required to transform one event-coded hybrid sequence into another through three edit operations: inserting, deleting, and replacing code terms. The single-operation cost for inserting and deleting code terms is fixed at 1.0, and the single-operation cost for replacing code terms is determined based on the semantic distance between the replaced code term and the replaced code term. Code terms are divided into two categories: event type codes and event source device codes, and the replacement cost calculation methods for these two types of codes are defined separately.

[0033] The replacement cost between event type codes is determined based on the hierarchical relationship of event classifications pre-stored in the production event classification hierarchy table. The production event classification hierarchy table is a tree-structured data table. The root node represents all events, and the first-level child nodes categorize events into five main categories: equipment events, manual events, handover events, material events, and alarm events. Each main category is further subdivided into intermediate and minor categories. The semantic similarity between two event type codes is determined by the hierarchical depth of their nearest common ancestor (LCA) node from the root node. The closer the LCA node is to a leaf node, the lower the replacement cost; the closer the LCA node is to the root node, the higher the replacement cost. For example, equipment start code E001 and equipment stop code E002 belong to the same subcategory, "Equipment Operation Status," under the equipment event category. Their LCA is the "Equipment Operation Status" node, and the replacement cost is set to 0.2. Equipment start code E001 and manual intervention code H011 belong to two different main categories: equipment events and manual events. Their LCA is the root node, and the replacement cost is set to 0.9. The replacement cost between the event source equipment codes is determined based on the upstream and downstream position distances of the equipment, which are pre-stored in the production line equipment topology table. The production line equipment topology table records the upstream and downstream adjacency relationships of all equipment in the supercritical foaming material production line, arranged according to the material flow direction. The material flow direction is, in order: melt pump unit, supercritical fluid injection unit, foaming vessel unit, and pressure relief control unit. The position distance between any two equipment is defined as the number of adjacent edges traversed by the shortest path in the topology table. The replacement cost equals the position distance divided by the maximum possible position distance in the topology table. For example, the position distance between melt pump unit MELT-PMP-05 and supercritical fluid injection unit SCF-INJ-03 is 1, and the maximum possible position distance is 3; the replacement cost is 1 divided by 3, approximately equal to 0.33. The position distance between melt pump unit MELT-PMP-05 and pressure relief control unit DEPR-CTL-12 is 3; the replacement cost is 3 divided by 3, equal to 1.0.

[0034] Based on the calculated weighted edit distance values ​​between all pairwise sequences, the risk mapping module constructs a hierarchical clustering tree structure. The construction of the hierarchical clustering tree structure employs a bottom-up agglomerative hierarchical clustering algorithm. Initially, each training sample data record is treated as an independent leaf node cluster. In each round of merging operations, the average weighted edit distance value between all pairwise clusters is calculated, and the pair with the smallest average weighted edit distance value is selected and merged into a new parent node cluster. The encoding sequence set of the parent node cluster is the union of the encoding sequence sets of its child node clusters. This merging operation is repeated until all clusters are merged into a single root node cluster, forming a complete binary tree-like hierarchical clustering tree structure.

[0035] After the hierarchical clustering tree structure is constructed, the risk mapping module traverses each cluster node layer by layer from the root node. For each cluster node encountered, the proportion of the complaint result identifier field in all training sample data records contained within that cluster node is calculated. During the calculation, the training sample data records within the cluster node are first grouped according to customer code, and the ratio between the number of records with a complaint result identifier field value of 1 in each customer code group and the total number of training sample data records in that customer code group is calculated. For example, if a cluster node contains 12 training sample data records for customer code CUST-08-2217, of which 3 are complaint records, then the ratio for this customer code group is 3 divided by 12, which equals 0.25. After calculating the ratios for all customer code groups, the arithmetic mean of the ratios for each customer code group is calculated. This arithmetic mean is the proportion of the complaint result identifier field for that cluster node. When the proportion of the complaint result identifier field for a cluster node exceeds a preset threshold of 0.6, the event-encoded mixed subsequence with the highest frequency in that cluster node is marked as a complaint-related feature subsequence. The most frequent event-coded mixed subsequences are extracted by scanning all training sample data records within the cluster node, counting the frequency of each continuous coded segment in the entire sequence, and selecting the most frequent continuous coded segment as the complaint-related feature subsequence. Each complaint-related feature subsequence is assigned a unique complaint risk propensity code, with a value ranging from R01 to R10. A higher code value indicates a higher complaint risk propensity for the corresponding complaint-related feature subsequence. Once all complaint-related feature subsequences are labeled, the production segment complaint risk mapping model is complete.

[0036] During the model application phase, the risk mapping module matches the event-coded mixed sequence of each production segment interval in the current production batch with all complaint-related feature subsequences stored in the production segment complaint risk mapping model. During matching, a weighted edit distance value is calculated between the event-coded mixed sequence of the current production segment interval and each complaint-related feature subsequence; the calculation method for the weighted edit distance value is the same as in the model building phase. The complaint-related feature subsequence with the smallest weighted edit distance value is selected as the matching result, and the complaint risk propensity code corresponding to this complaint-related feature subsequence is determined as the output of that production segment interval. If multiple complaint-related feature subsequences have the same weighted edit distance value, the one with the higher complaint risk propensity code value is selected. Thus, each production segment interval is assigned a specific complaint risk propensity code for subsequent use by the customer order allocation rule base.

[0037] In another preferred embodiment of the present invention, the specific process of outputting the allocation binding relationship record after performing the matching operation in the matching allocation module is as follows: The matching and allocation module first obtains the complaint event frequency distribution vectors for all customer codes within the historical time period from the vector extraction module. The complaint event frequency distribution vector is a multi-dimensional numerical array, with the number of dimensions equal to the total number of deduplicated complaint issue type codes. Taking customer code CUST-08-2217 as an example, its complaint event frequency distribution vector has a value of 3.27 in dimension Q001, 1.18 in dimension Q003, 0.93 in dimension Q004, and 0 in the remaining dimensions. This vector is stored as an array [3.27, 0, 1.18, 0.93, 0, 0, 0], with the array indices corresponding one-to-one with the ascending order of the complaint issue type code dimensions.

[0038] The matching and assignment module calculates the cosine distance between the complaint event frequency distribution vectors of every two customer codes. The cosine distance is defined as 1 minus the cosine of the angle between the two vectors. The cosine of the angle between two vectors A and B is equal to the inner product of A and B divided by the product of the magnitudes of A and B. The cosine distance ranges from 0 to 2; a smaller value indicates that the two vectors are closer in direction, meaning the two customers have more similar complaint issue type distribution structures. For example, the cosine distance between the inner product of the vectors of customer code CUST-08-2217 and customer code CUST-11-0532 is 0.23, while the cosine distance between the vector of CUST-08-2217 and customer code CUST-11-0532 is 1.47. The matching and assignment module sets a preset initial clustering radius of 0.35, grouping any two customer codes with a cosine distance less than 0.35 into the same customer group. The clustering process uses a transitive closure approach, meaning that if the cosine distance between customer A and customer B is less than 0.35, and the cosine distance between customer B and customer C is less than 0.35, then customers A, B, and C are all grouped into the same customer group. Each customer group is assigned a group identifier code, such as GRP-01, GRP-02, and GRP-03. Each customer code uniquely belongs to one customer group.

[0039] A corresponding risk tolerance threshold range is set for each customer group. The risk tolerance threshold range is defined as a closed interval, with the lower limit being the lower tolerance limit and the upper limit being the upper tolerance limit. Both limits are taken from the coded value range R01 to R10 of the complaint risk propensity code. Complaint risk propensity code R01 corresponds to the lowest complaint risk propensity, and R10 corresponds to the highest complaint risk propensity. The initial setting method for the risk tolerance threshold range is to take the minimum value of the complaint risk propensity code among all customer codes within the customer group in the historical delivery records of the production segment interval actually received as the lower tolerance limit, and the maximum value as the upper tolerance limit. If the number of customers in the customer group is less than 3, the lower tolerance limit is set to R01, and the upper tolerance limit is set to R10. For example, the initial risk tolerance threshold range for customer group GRP-01 is R02 to R06, and the initial risk tolerance threshold range for customer group GRP-02 is R04 to R09.

[0040] After each batch of supercritical foamed material production is delivered and complaint information is received from each customer code, the matching and allocation module initiates a dynamic adjustment process for customer groups and threshold ranges. For each customer code, the segment numbers of all production segment ranges allocated to that customer code in that production batch are obtained, and the complaint risk tendency codes corresponding to these segment numbers are retrieved from the risk mapping module. The complaint risk tendency code closest to R10 is extracted, i.e., the code value with the highest complaint risk tendency, and is recorded as the highest risk code for this delivered segment. This highest risk code for the segment is compared with the risk tolerance threshold range of the customer group to which the customer code belongs. The comparison rule is to determine whether the code sequence number of the highest risk code for the segment is greater than the code sequence number of the upper tolerance limit. If it is greater, it is determined to be outside the risk tolerance threshold range; if it is less than or equal to, it is determined to be within the risk tolerance threshold range.

[0041] If the highest-risk code of a segment exceeds the risk tolerance threshold range and no complaint is filed for that customer code after this delivery, the positive adjustment cumulative value associated with that customer code will be increased by 1. The initial value of the positive adjustment cumulative value is 0, and each customer code maintains an independent positive adjustment cumulative value counter. If the highest-risk code of a segment falls within the risk tolerance threshold range and a complaint is filed for that customer code after this delivery, the negative adjustment cumulative value associated with that customer code will be increased by 1. The initial value of the negative adjustment cumulative value is 0, and each customer code maintains an independent negative adjustment cumulative value counter. The preset positive adjustment threshold is set to 3 times, and the preset negative adjustment threshold is set to 2 times.

[0042] When the cumulative positive adjustment value of a customer code exceeds a preset positive adjustment threshold of 3, the matching and allocation module removes the customer code from its current customer group and calculates the cosine distance between the customer code's complaint event frequency distribution vector and the group center vectors of all remaining customer groups. The group center vector is defined as the arithmetic mean vector of the complaint event frequency distribution vectors of all customer codes within that group. The removed customer code is then reassigned to the customer group with the smallest cosine distance, and its cumulative positive adjustment value is reset to 0. When the cumulative negative adjustment value of any customer code within a customer group exceeds a preset negative adjustment threshold of 2, the matching and allocation module shrinks the upper and lower boundary values ​​of the risk tolerance threshold range for that customer group. The shrinkage method involves subtracting 1 from the code number of the upper tolerance limit and adding 1 to the code number of the lower tolerance limit. If the upper tolerance limit is less than the lower tolerance limit after shrinkage, they are kept equal. After shrinkage, the cumulative negative adjustment values ​​of all customer codes within that customer group are reset to 0.

[0043] When performing the matching operation between the production segment interval to be assigned and the customer code to be assigned, the matching and allocation module first obtains the complaint risk tendency code of the current production segment interval to be assigned from the risk mapping module, and then determines the customer group to which the customer code to be assigned belongs and the corresponding risk tolerance threshold range. The code sequence number of the complaint risk tendency code is compared with the code sequence numbers of the lower tolerance limit and the upper tolerance limit, respectively. If the code sequence number of the complaint risk tendency code is greater than or equal to the code sequence number of the lower tolerance limit and less than or equal to the code sequence number of the upper tolerance limit, it is determined that it falls within the risk tolerance threshold range. At this time, the matching and allocation module outputs an allocation binding relationship record, which contains the segment number of the current production segment interval and the customer code to be assigned. If the complaint risk tendency code does not fall within the risk tolerance threshold range, no allocation binding relationship record is output for that customer code, and the matching operation continues with the next customer code to be assigned.

[0044] In another preferred embodiment of the present invention, the specific process of attaching a complaint risk propensity code to the delivery document in the outbound identification module is as follows: The outbound identification module performs the association storage of finished product packaging box identification codes and customer codes, as well as the processing of additional information on shipping documents, during the finished product outbound business processing phase of the Enterprise Resource Planning (ERP) system. After the matching and allocation module completes the output of allocation and binding relationship records for all production segment intervals of the current production batch, the outbound identification module writes these allocation and binding relationship records into the outbound shipment details table. The outbound shipment details table is a structured data table containing the following fields: finished product packaging box identification code, production batch number, production segment interval number, customer code, complaint risk propensity code, shipment date, and carrier code. Each finished product packaging box identification code uniquely corresponds to a production segment interval number, thereby indirectly associating it with a customer code and a complaint risk propensity code.

[0045] During the finished product packaging outbound scanning process, warehouse staff use a barcode scanner to read the packaging identification barcode on the surface of the finished product packaging boxes. The scanner sends the read packaging identification code to the warehouse execution terminal. The warehouse execution terminal retrieves the corresponding customer code and complaint risk propensity code from the outbound shipment details table based on the packaging identification code, and then calls the label printing command to generate and output a logistics label on the thermal transfer logistics label printer. The printed content of the logistics label includes the recipient's address information, the shipper's address information, the carrier's barcode, and a two-dimensional barcode graphic. The encoded content of the two-dimensional barcode graphic is a Uniform Resource Locator (URL) string, with the format: https: / / erp.companydomain.com / acceptance_prompt?customer_code=customer_code_value&risk_code=complaint_risk_propensity_code_value. For example, https: / / erp.companydomain.com / acceptance_prompt?customer_code=CUST-08-2217&risk_code=R05.

[0046] In the logistics process, the receiving party's unloading personnel use handheld scanning terminals to scan the two-dimensional barcode images on the logistics labels. The handheld scanning terminal parses the Uniform Resource Locator (URI) string in the two-dimensional barcode image, extracting the customer code parameter value and risk code parameter value. Simultaneously, the handheld scanning terminal obtains its current geographical location information, including longitude and latitude coordinates, through its built-in satellite positioning module. The handheld scanning terminal assembles the customer code parameter value, risk code parameter value, longitude coordinate value, and latitude coordinate value into a query request data packet, which is then sent to the query interface of the Enterprise Resource Planning (ERP) system via a wireless network.

[0047] After receiving a query request data packet, the query interface of the Enterprise Resource Planning (ERP) system first retrieves the frequency distribution of the corresponding historical complaint issue types based on the risk coding parameter value. The retrieval method involves scanning all complaint records in the historical complaint database where the complaint risk tendency coding field value is the specified risk coding parameter value, counting the frequency of each complaint issue type code, sorting them by frequency from highest to lowest, and taking the top three complaint issue type codes and their corresponding frequencies. For example, in the historical complaint records corresponding to risk code R05, Q001 surface defects appeared 12 times, Q003 substandard mechanical properties appeared 5 times, and Q004 uneven foaming ratio appeared 3 times.

[0048] The Enterprise Resource Planning (ERP) system simultaneously retrieves the number of historical cargo damage records for transit warehouses along the logistics route within a historical time period, based on both longitude and latitude coordinates. The search method first identifies the logistics transit warehouse with the closest straight-line distance to the current latitude and longitude coordinates, calculated using the spherical cosine theorem. Then, using the warehouse code of that transit warehouse as the query condition, it retrieves the total number of cargo damage records associated with that transit warehouse in the cargo damage record database over the past 180 days. For example, if the nearest transit warehouse is the Zhengzhou transit warehouse (warehouse code ZZH-01), there are 8 historical cargo damage records in the past 180 days.

[0049] The Enterprise Resource Planning (ERP) system inputs the frequency of complaint types and the number of historical damage records into a preset acceptance prompt text generation template. The acceptance prompt text generation template is a string concatenation rule combining a set of fixed text fragments and variable parameters. The template content is: "Common complaint types for this batch of products in historical batches of the same risk level are: Type 1 name appears once and several times; Type 2 name appears twice and several times; Type 3 name appears twice and several times. The nearest transit warehouse along this transportation route has a total of [number] damage records in the past six months. It is recommended to focus on checking the product appearance and the items corresponding to the above complaint types during unloading." The "Type 1 name" and "Type 2 frequency" are variable parameters filled in from the search results. For example, the filled unloading acceptance prompt text would be: "Common complaint types for this batch of products in historical batches of the same risk level are: surface defects appeared 12 times, mechanical performance failures appeared 5 times, and uneven foaming ratio appeared 3 times. The nearest transit warehouse along this transportation route has a total of 8 damage records in the past six months. It is recommended to focus on checking the product appearance and the items corresponding to the above complaint types during unloading." The Enterprise Resource Planning (ERP) system returns the generated unloading and acceptance prompt text as response data to the handheld scanning terminal, which then displays the text on its screen.

[0050] 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. An AI-enabled monitoring and management system for anomalies in the production of supercritical foamed materials, characterized in that, include: The log input module is used to input a non-parametric production event log file for the production batch of supercritical foamed materials. The non-parametric production event log file contains production event entries recorded in timestamp order. The vector extraction module is used to retrieve the product sales flow database based on the batch number of the production batch, obtain the customer code list, retrieve the customer historical complaint record database for each customer code, and extract the frequency distribution vector of complaint events. The segment encoding module is used to divide production event entries into multiple production segment intervals according to the production scheduling timeline, extract the event type encoding sequence within each production segment interval, and interleave the event type encoding field and the event source equipment encoding field according to the chronological order of the event occurrence timestamps of their respective production event entries to form a mixed event encoding sequence; The risk mapping module is used to input the event type coding sequence of each production segment interval into the production segment complaint risk mapping model. The production segment complaint risk mapping model is constructed based on the co-occurrence relationship between historical event coding combinations and complaint records, and outputs a complaint risk tendency code. The process includes: obtaining event-coded mixed sequences for each historical production segment interval in the historical production batch set and combining them with customer complaint result identifier fields to form training sample data records; calculating the weighted edit distance between every two event-coded mixed sequences in the training sample data record set; constructing a hierarchical clustering tree structure based on the weighted edit distance values; statistically analyzing the proportion of the complaint result identifier field at each cluster node in the hierarchical clustering tree structure; when the proportion exceeds a preset proportion threshold, marking the event-coded mixed subsequence with the highest frequency in the cluster node as a complaint-related feature subsequence and assigning it a corresponding complaint risk tendency code; matching the event-coded mixed sequence of the current production segment interval with each complaint-related feature subsequence; and determining the complaint risk tendency code corresponding to the complaint-related feature subsequence with the highest matching degree as the output. The matching and allocation module is used to input the frequency distribution vector of complaint events for each customer code and the complaint risk tendency code for each production segment into the customer order allocation rule base, and output the allocation binding relationship record after performing the matching operation. This includes: clustering customer codes based on the cosine distance of the frequency distribution vector of complaint events for each customer code within a historical time period and setting a risk tolerance threshold range for each customer group; after each production batch is delivered, obtaining the complaint risk tendency code for the production segment interval assigned to the customer code; if the complaint risk tendency code exceeds the risk tolerance threshold range and the customer code has not complained, recording a positive adjustment cumulative value; if the complaint risk tendency code falls within the risk tolerance threshold range and the customer code complains, recording a negative adjustment cumulative value; when the positive adjustment cumulative value exceeds a preset positive adjustment threshold, re-determining the customer group to which the customer code belongs; when the negative adjustment cumulative value exceeds a preset negative adjustment threshold, shrinking the upper and lower boundary values ​​of the risk tolerance threshold range of the customer group; and when performing a matching operation, comparing the complaint risk tendency code of the current production segment interval with the risk tolerance threshold range of the customer group to which the customer code to be assigned belongs, and outputting the allocation binding relationship record if they fall within the range. The outbound identification module is used to associate the finished product packaging box identification code with the customer code in the outbound shipment details table during the finished product outbound stage of the enterprise resource planning system, based on the allocation binding relationship record, and to attach a complaint risk tendency code to the shipment document.

2. The AI-enabled supercritical foaming material production anomaly monitoring and management system according to claim 1, characterized in that, In the log input module, the specific process of inputting the non-parametric production event log file for the supercritical foaming material production batch is as follows: The non-parametric production event log file is generated by the production execution system one by one according to the event trigger time during the operation of the supercritical foaming material production line. The production event entries include equipment start-up and shutdown event entries, manual intervention operation event entries, production shift handover event entries, material feeding switching event entries, and equipment alarm reset event entries. Each production event entry also includes an event source equipment code field and an event supplementary description text field. The event supplementary description text field is written by obtaining the event type code fields of all production event entries within a preset time window before the event occurrence timestamp and concatenating them in reverse chronological order to form an event prefix code string. Then, the event type code fields of all production event entries within a preset time window after the event occurrence timestamp are obtained and concatenated in ascending chronological order to form an event suffix code string. The event prefix code string and the event suffix code string are connected with a preset separator and then written into the event supplementary description text field.

3. The AI-enabled supercritical foaming material production anomaly monitoring and management system according to claim 1, characterized in that, In the vector extraction module, the specific process of retrieving the customer's historical complaint record database and extracting the frequency distribution vector of complaint events for each customer code is as follows: Retrieve all complaint records generated by the customer code within the historical time period from the customer history complaint record database. Each complaint record contains a complaint issue type code field and a complaint occurrence date field. After deduplication of all complaint issue type codes that occurred within the historical time period, sort them in ascending order by code value to form a complaint issue type code dimension list. For each complaint issue type code in the complaint issue type code dimension list, obtain the set of complaint records corresponding to the complaint issue type code of the customer code within the historical time period. Assign a time decay weight value to each complaint record in the complaint record set. The time decay weight value is calculated by calculating the date difference between the complaint occurrence date field and the base date using the current system date as the base date. Input the date difference into a preset negative exponential decay function to obtain the time decay weight value of the complaint record. Accumulate the time decay weight values ​​of all complaint records in the complaint record set, and determine the accumulated result as the vector component value corresponding to the dimension in the complaint event frequency distribution vector.

4. The AI-enabled supercritical foaming material production anomaly monitoring and management system according to claim 1, characterized in that, In the segment encoding module, the specific process of dividing production event entries into multiple production segment intervals according to the production scheduling timeline and extracting the event type encoding sequence within each production segment interval is as follows: Obtain the event occurrence timestamps of all production shift handover event entries recorded on the production scheduling timeline for the production batch. Determine the time span between the event occurrence timestamps of two adjacent production shift handover event entries as the duration of a production segment interval. Each production segment interval corresponds to a segment number. Extract the event type code field and the event source equipment code field of all production event entries occurring within each production segment interval. Arrange the event type code field and the event source equipment code field in an alternating manner according to the chronological order of the event occurrence timestamps of their respective corresponding production event entries to form an event code hybrid sequence. Determine the event code hybrid sequence as the event type code sequence corresponding to the production segment interval.

5. The AI-enabled supercritical foaming material production anomaly monitoring and management system according to claim 1, characterized in that, The weighted edit distance value is calculated as follows: the replacement cost between event type codes is determined based on the event classification hierarchy relationship stored in the production event classification hierarchy table; the replacement cost between event source equipment codes is determined based on the upstream and downstream location distance of the equipment stored in the production line equipment topology relationship table; and the proportion value of the statistical complaint result identifier field is calculated by first grouping the training sample data records in the cluster node according to customer code, calculating the ratio of the number of complaint records in each customer code group to the total number of training sample data records, and then calculating the arithmetic mean of the ratios of each customer code group.

6. The AI-enabled supercritical foaming material production anomaly monitoring and management system according to claim 1, characterized in that, In the outbound identification module, the specific process of attaching a complaint risk tendency code to the delivery document is as follows: The warehouse execution system prints a two-dimensional barcode graphic containing the customer code and complaint risk propensity code on the logistics label based on the correspondence between the finished product packaging box identification code and the customer code in the outbound shipment details table. When the handheld scanning terminal reads the two-dimensional barcode graphic, it sends a query request containing the complaint risk propensity code and the current location information to the enterprise resource planning system. The enterprise resource planning system retrieves the historical complaint problem type distribution frequency based on the complaint risk propensity code, and retrieves the historical cargo damage record number through transit warehouses based on the current location information. It then inputs the historical complaint problem type distribution frequency and the historical cargo damage record number into the acceptance prompt text generation template to generate an unloading acceptance prompt text, which is then returned to the handheld scanning terminal for display.

Images