Aluminum alloy die casting quality prediction system and method

Abstract

The invention belongs to the technical field of production management, and provides an aluminum alloy die casting quality prediction system and method. The method comprises the following steps: obtaining a causal association comparison table based on quality defect record data of the aluminum alloy die casting; comparing each defect prediction type of the current round with a causal association comparison table to obtain a plurality of key parameter types corresponding to any defect prediction type; collecting and selecting data corresponding to each key parameter type from the real-time die-casting parameters to form prediction basic data; and performing prediction processing on the prediction basic data by using the quality defect prediction model corresponding to the defect prediction type to obtain a defect probability of the aluminum alloy die casting corresponding to the defect prediction type, and outputting an alarm signal when the defect probability is higher than a probability threshold. Compared with a traditional experience judgment method, scientific selection of key die-casting parameters is achieved, the accuracy and timeliness of quality defect prediction are effectively improved, the rejection rate is reduced, and the production efficiency and the product quality stability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of production management, and in particular to a quality prediction method and system for aluminum alloy die castings. Background Art

[0002] In modern manufacturing, aluminum alloys, with their significant advantages such as low density, high specific strength, excellent corrosion resistance, and superior casting properties, have become one of the preferred materials for die-casting in numerous industries, including automotive, aerospace, and electronic communications. With the rapid development of autonomous driving technology, aluminum alloy castings are widely used in autonomous driving camera modules due to their excellent overall performance. As a core component of a vehicle's visual perception system, autonomous driving camera modules require extremely high precision, stability, and reliability of their structural components. Aluminum alloy castings can meet the module's lightweight requirements and reduce vehicle load. Their excellent heat dissipation properties also help the camera operate stably in complex environments, ensuring the clarity and accuracy of image acquisition.

[0003] However, the aluminum alloy die-casting process is a complex one involving the coupling of multiple physical fields, including material flow, heat transfer, and solidification. The quality of aluminum alloy castings used in unmanned camera modules is affected by numerous factors, including mold design, die-casting process parameters (such as die-casting temperature, die-casting pressure, and filling speed), raw material properties, and equipment status. These factors can easily lead to defects such as pores, shrinkage, cracks, and cold shuts. Once quality issues arise, not only will the assembly accuracy of the camera module decrease, affecting the image acquisition angle and stability, but they can also cause poor heat dissipation, leading to camera failures such as blurred images and reduced frame rates due to overheating. These issues directly threaten the safety of the unmanned vehicle's perception system and, in turn, driving safety. This results in an extremely low tolerance for scrap rates for aluminum alloy castings in this field, significantly increasing the difficulty of quality control.

[0004] Currently, quality control for aluminum alloy die-castings primarily relies on traditional empirical judgment and post-testing methods. Empirical judgment relies on the operator's long-term accumulated production experience, is highly subjective, and is difficult to standardize. Different operators have varying judgment criteria and operating habits, resulting in poor product quality stability. Post-testing, on the other hand, involves inspecting die-castings after production is complete through visual inspection and non-destructive testing (such as X-ray and ultrasonic testing). While this method can identify quality issues, it prevents early intervention during the production process. Once a quality issue is discovered, it wastes raw materials, energy, and time, increasing production costs and reducing efficiency.

[0005] Therefore, there is an urgent need for a more scientific, accurate and reliable quality prediction method for aluminum alloy die castings to improve product quality and production efficiency and reduce production costs. Summary of the Invention

[0006] To this end, the present invention provides a method, system, electronic device, computer storage medium and computer program product for predicting the quality of aluminum alloy die castings to solve at least one of the above technical problems.

[0007] In a first aspect, the present invention provides a method for predicting the quality of aluminum alloy die-castings, comprising the following method steps: collecting quality defect record data of aluminum alloy die-castings, performing causal analysis on each quality defect record data, and obtaining a causal comparison table; the quality defect record data includes quality defect types and corresponding historical die-casting parameters; determining each defect prediction type of this round, comparing each defect prediction type with the causal comparison table, and obtaining several key parameter types corresponding to any defect prediction type; collecting real-time die-casting parameters in the aluminum alloy die-casting process, and selecting data corresponding to each of the key parameter types to constitute prediction basic data; using the quality defect prediction model corresponding to the defect prediction type to perform prediction processing on the prediction basic data, and obtaining the defect probability of the aluminum alloy die-casting corresponding to the defect prediction type, and outputting an alarm signal when the defect probability is higher than the probability threshold.

[0008] According to a second aspect of the present invention, a quality prediction system for aluminum alloy die-castings is provided, the system comprising a causal correlation combing module, a key parameter determination module, a die-casting parameter acquisition module, and a quality defect prediction module; the causal correlation combing module collects quality defect record data of aluminum alloy die-castings, performs causal correlation analysis on each quality defect record data, and obtains a causal correlation comparison table; the quality defect record data includes quality defect types and corresponding historical die-casting parameters; the key parameter determination module determines each defect prediction type of this round, compares each defect prediction type with the causal correlation comparison table, and obtains several key parameter types corresponding to any defect prediction type; the die-casting parameter acquisition module collects real-time die-casting parameters in the aluminum alloy die-casting process, and selects data corresponding to each key parameter type therefrom to constitute prediction basic data; the quality defect prediction module uses the quality defect prediction model corresponding to the defect prediction type to perform prediction processing on the prediction basic data, obtains the defect probability of the aluminum alloy die-casting corresponding to the defect prediction type, and outputs an alarm signal when the defect probability is higher than the probability threshold.

[0009] According to a third aspect of the present invention, an electronic device is provided, comprising: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program implements any of the methods described above when executed by the processor.

[0010] According to a fourth aspect of the present invention, a computer storage medium is provided, wherein the computer storage medium stores a computer program executable by a processor to implement any of the methods described above.

[0011] According to a fifth aspect of the present invention, a computer program product is provided, which comprises a computer program executable by a processor to implement any of the methods described above.

[0012] This method collects quality defect records and uses correlation analysis to generate a causal correlation table, accurately matching defects with key parameters. Furthermore, it collects and filters die-casting parameters in real time to form prediction data. Defect probabilities are calculated using corresponding quality defect prediction models, and alarms are triggered when thresholds are exceeded. Compared to traditional empirical judgment methods, this method achieves scientific selection of key die-casting parameters, effectively improving the accuracy and timeliness of quality defect prediction, reducing scrap rates, and increasing production efficiency and product quality stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0014] Figure 1 It is a flow chart of a method for predicting the quality of aluminum alloy die castings disclosed in an embodiment of the present invention.

[0015] Figure 2 It is a structural diagram of the quality defect prediction model disclosed in an embodiment of the present invention.

[0016] Figure 3 It is a structural schematic diagram of an aluminum alloy die casting quality prediction system disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The following specific embodiments illustrate the implementation of this application. Those familiar with the art can easily understand the other advantages and functions of this application from the content disclosed in this specification. Obviously, the described embodiments are part of the embodiments of this application, but not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0018] In addition, the technical features involved in the different embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0019] like Figure 1 As shown, an embodiment of the present invention discloses a method for predicting the quality of aluminum alloy die-castings, including the following method steps: S10, collecting quality defect record data of aluminum alloy die-castings, performing causal correlation analysis on each quality defect record data, and obtaining a causal correlation comparison table; the quality defect record data includes quality defect types and corresponding historical die-casting parameters.

[0020] In the field of aluminum alloy die-casting, existing technologies have long relied on operator experience to determine the key die-casting parameters corresponding to various defect types. This approach has significant limitations. Empirical judgments lack a quantitative basis, and different operators' operating habits and judgment criteria vary greatly. This leads to highly subjective and inaccurate selection of key parameters, making it difficult to effectively address the complex multi-factor coupling in the die-casting process. Consequently, it is impossible to accurately predict quality defects such as porosity, shrinkage, and cracks, resulting in high scrap rates and a serious constraint on improving production efficiency and product quality.

[0021] To address these issues, the present invention systematically establishes the intrinsic links between key die-casting parameters and quality defects through data-driven causal analysis. Specifically, a large amount of quality defect data, including quality defect types (such as pores, shrinkage, and cracks) and corresponding historical die-casting parameters (such as pouring temperature, injection pressure, and filling speed), is collected. The causal relationships are then deeply explored to form a causal correlation table. This transforms previously vague empirical data into objective, quantified parameter-defect correspondences, known as the causal correlation table.

[0022] Causal analysis, based on algorithms such as Apriori and Bayesian networks, can deeply mine large amounts of quality defect records. By calculating the correlation and confidence between different parameter combinations and the occurrence of quality defects, it can identify which die-casting parameter changes significantly affect the occurrence of specific defects. For example, it was found that when the pouring temperature exceeds a certain threshold and the holding time is insufficient, the probability of shrinkage defects increases sharply, thus revealing the causal relationship between parameters and defects.

[0023] The causal correlation table can be referred to Table 1 below: Table 1:

[0024] It should be noted that the above Table 1 of the present invention is only used for illustrative purposes, and is not intended to limit the cause-and-effect correlation table in the present invention to be based only on the above Table 1.

[0025] S20, determining each defect prediction type of this round, comparing each defect prediction type with the causal association comparison table, and obtaining several key parameter types corresponding to any defect prediction type.

[0026] The quality defect types to be prioritized for prediction in this round are determined based on production requirements, product characteristics, or historical defect prevalence. Specifically, for aluminum alloy die-castings used in autonomous driving camera modules, defects such as pores and deformation, which have a significant impact on imaging accuracy, are prioritized as prediction targets.

[0027] The causal correlation table accurately records the key parameter types corresponding to each defect type. By comparing the selected defect prediction type with the causal correlation table, the die-casting parameters closely related to the currently predicted defect can be quickly identified. For example, when predicting porosity defects, the table directly links to parameters such as pouring temperature, mold temperature, filling speed, and alloy liquid purity. This eliminates subjective guesswork based on experience and significantly improves the accuracy and efficiency of parameter selection.

[0028] S30, collecting real-time die-casting parameters in the aluminum alloy die-casting process, and selecting data corresponding to each key parameter type to form prediction basic data.

[0029] Using temperature sensors, pressure sensors, speed sensors, and other equipment, we continuously monitor various parameters during the die-casting process in real time, acquiring real-time die-casting parameters encompassing temperature, pressure, speed, time, and other dimensions, ensuring coverage of all parameters that may impact casting quality. These parameters include, but are not limited to, pouring temperature, mold temperature, low-speed shot pressure, high-speed shot pressure, holding pressure, low-speed shot speed, high-speed shot speed, filling time, holding time, mold cooling time, alloy liquid purity data, and mold exhaust system status data.

[0030] Based on the key parameter type corresponding to the defect prediction type determined in step S20, the corresponding key data is accurately screened from the real-time collected die-casting parameters. For example, if a defect prediction type is associated with the pouring temperature and injection speed parameters, only the real-time data of these two parameters is extracted, and other irrelevant parameters are eliminated to form the core data set for quality prediction.

[0031] The on-demand screening mechanism of the present invention can effectively reduce data redundancy, focus on key information, and provide high-quality input data for subsequent quality defect prediction models.

[0032] S40, using the quality defect prediction model corresponding to the defect prediction type to perform prediction processing on the prediction basic data, to obtain the defect probability of the aluminum alloy die casting corresponding to the defect prediction type, and output an alarm signal when the defect probability is higher than a probability threshold.

[0033] Pre-train corresponding quality defect prediction models (e.g., machine learning models such as neural networks and random forests) for different defect prediction types. The prediction basic data generated in step S30 is input into the corresponding quality defect prediction model. Based on the quantitative relationship between parameters and defects learned during the training phase, the quality defect prediction model analyzes and calculates real-time data and outputs the probability of that type of defect occurring in aluminum alloy die-castings. For example, if the input is prediction basic data related to porosity defects, the quality defect prediction model can calculate the probability of porosity defects occurring under current production conditions.

[0034] A reasonable probability threshold is set for each defect prediction type. When the defect probability output by the quality defect prediction model exceeds this probability threshold, an alarm mechanism is immediately triggered. For example, an alarm is issued to production personnel through audio and visual prompts, message push, etc. At the same time, preset rules can be combined to provide die-casting parameter adjustment suggestions, helping personnel to intervene in the production process in a timely manner. This achieves a closed-loop management from defect prediction to proactive control, making up for the lack of early prevention in traditional post-detection.

[0035] This method collects quality defect records and uses correlation analysis to generate a causal correlation table, accurately matching defects with key parameters. Furthermore, it collects and filters die-casting parameters in real time to form prediction data. Defect probabilities are calculated using corresponding quality defect prediction models, and alarms are triggered when thresholds are exceeded. Compared to traditional empirical judgment methods, this method achieves scientific selection of key die-casting parameters, effectively improving the accuracy and timeliness of quality defect prediction, reducing scrap rates, and increasing production efficiency and product quality stability.

[0036] As an example, determining the defect prediction types of this round includes: retrieving several trial production data corresponding to this round, and the quality inspection results corresponding to each trial production data, and determining each defect type in the quality inspection results corresponding to the last trial production as the defect prediction type.

[0037] In actual production, multiple trial runs of aluminum alloy die-castings are conducted during a new round of production. After each trial run, samples are quality inspected to determine the defect types corresponding to that trial run. Die-casting equipment is regulated based on the defect types, and then trial runs, quality inspections, and regulation are conducted again. Thus, the defect types in the quality inspection results corresponding to the last trial run are defects that cannot be completely avoided by regulating the die-casting equipment. These defects are identified as the defect types that require in-depth prediction using the quality defect prediction model in this round.

[0038] The method of the present invention, which is based on the quality inspection feedback results of actual trial production, can focus on quality problems exposed recently, make subsequent predictions more in line with production reality, ensure the accuracy of the prediction direction, and effectively improve the prediction pertinence and production quality control efficiency.

[0039] As an example, the causal association analysis of each quality defect record data to obtain a causal association comparison table includes: using the Apriori algorithm to perform preliminary association rule mining on each quality defect record data to generate a first association rule set; each first association rule in the first association rule set has a corresponding first association strength; using the Bayesian network algorithm to perform causal structure learning on each quality defect record data, construct a probabilistic dependency relationship between die-casting parameters and defect types, and generate a second association rule set; each second association rule in the second association rule set has a corresponding second association strength; cross-validating and fusing the first association rule set and the second association rule set to retain strong association rules supported by both algorithms; using the Bayesian network algorithm to verify the first association rules in the first association rule set that do not belong to the strong association rules and whose first association strength is higher than a threshold, and using the Apriori algorithm to verify the second association rules in the second association rule set that do not belong to the strong association rules and whose second association strength is higher than a threshold, to obtain several first association rules and second association rules that meet the verification conditions, and combining them with the strong association rules to construct the causal association comparison table.

[0040] In order to improve the accuracy of constructing the causal association comparison table, the present invention combines the Apriori algorithm with the Bayesian network algorithm.

[0041] First, the Apriori algorithm is used to process the quality defect record data, mining preliminary association rules to form a first association rule set. Each first association rule is assigned a corresponding first association strength to measure the credibility of the first association rule. The specific processing process is as follows: 1. Data preprocessing: The quality defect record data is converted into a transaction database format. Each transaction contains a set of parameter values ​​and the corresponding defect type.

[0042] For example: [pouring temperature = 680℃, mold temperature = 220℃, filling speed = 4.2m / s, defect type = pore].

[0043] [Pouring temperature = 710℃, mold temperature = 250℃, filling speed = 3.8m / s, defect type = shrinkage].

[0044] Divide continuous parameters (such as temperature and speed) into intervals (such as pouring temperature ∈ [670,700)) to facilitate the generation of Boolean item sets.

[0045] 2. Frequent item set generation: Generate frequent item sets through layer-by-layer search (k-item set → k+1-item set). The specific steps are as follows: (1) Set the minimum support threshold (for example, min_support = 0.2, indicating that the item set must appear in more than 20% of the transactions).

[0046] Generate all individual items (e.g., {pouring temperature = 680°C}, {defect type = pore}) and calculate their support.

[0047] (2) Combine the frequent k-items into a candidate k+1-item set.

[0048] For example: if {pouring temperature>700℃, filling speed>4m / s} and {defect type=pore} are both frequent item sets, then the candidate {pouring temperature>700℃, filling speed>4m / s, defect type=pore} is generated.

[0049] Eliminate candidate sets that contain infrequent k-itemsets.

[0050] Scan the database, count the support of each candidate set, and retain the item sets that meet min_support.

[0051] Stop when no larger frequent itemsets can be generated.

[0052] 3. Association rule generation: (1) Extract association rules that meet the minimum confidence from frequent item sets. The steps are as follows: For each frequent item set X, generate all possible non-empty subsets A⊂X and construct the rule A→(XA).

[0053] The frequent item set {pouring temperature > 700℃, filling speed > 4m / s, defect type = pore} can generate the following rules: {pouring temperature > 700℃, filling speed > 4m / s} → {defect type = pore}; {pouring temperature > 700℃} → {filling speed > 4m / s, defect type = pore}.

[0054] (2) The confidence of rule A→B is defined as: .

[0055] For example, if the support for {pouring temperature > 700°C, filling speed > 4 m / s, defect type = pore} is 0.25, and the support for {pouring temperature > 700°C, filling speed > 4 m / s} is 0.35, then the rule confidence is 0.25 / 0.35≈0.714.

[0056] (3) The rules with confidence ≥ the minimum confidence threshold (e.g., min_confidence = 0.6) are retained to form the first association rule set.

[0057] 4. Quantifying Association Strength: Each first association rule is assigned two association strength metrics: Support: This indicates the universality of the rule, that is, the probability that an item set (e.g., "die casting temperature is too high" and "casting cracks") will co-occur in all data: support(A→B)=P(A∪B). For example, if 30 out of 100 records contain both "temperature is too high" and "casting cracks," the support is 30%.

[0058] Confidence: This indicates the reliability of a rule, that is, the probability of B occurring given the presence of A: confidence(A→B)=P(B|A). For example, if there are 50 records containing "overtemperature," and 30 of them also contain "casting cracks," the confidence is 30 / 50=60%.

[0059] Lift: Measures whether the correlation between A and B is greater than that of random independence. It indicates the degree to which the occurrence of the antecedent A increases the probability of the occurrence of the consequent B. The calculation formula is: lift(A→B) = confidence(A→B) / P(B). Lift > 1 indicates a positive correlation (for example, lift = 1.8 means the probability of A and B occurring together is 1.8 times higher than in the random case); lift = 1 indicates independence; lift < 1 indicates a negative correlation. For example, if the overall support for "casting cracks" is 20%, the lift in the above example is 60% / 20% = 3, indicating that "excessive temperature" significantly increases the probability of "casting cracks."

[0060] The first correlation strength is calculated based on support, confidence, and lift. For example, a weighted comprehensive scoring method can be used, with the calculation formula being: a1×Support+a2×Confidence+a3×Lift. a1-a3 are weight coefficients, and the sum of the weight coefficients is 1. Alternatively, the geometric mean method can be used, with the calculation formula being: The present invention does not impose any specific limitation on this.

[0061] An example of the first association rule set obtained through steps 1-4 above is shown in Table 2 below: Table 2:

[0062] It should be noted that the above Table 2 of the present invention is only used for illustrative purposes, and is not intended to limit the first association rule set in the present invention to only include the contents of the above Table 2.

[0063] Next, the Bayesian network algorithm is used to learn the causal structure of each quality defect record data, construct the probabilistic dependency relationship between die-casting parameters and defect types, and generate a second association rule set. Similarly, each second association rule is assigned a corresponding second association strength to reflect the degree of probabilistic association between die-casting parameters and defects. The specific processing process is as follows: 1. Data preprocessing and discretization: Continuous die-casting parameters (such as pouring temperature and filling speed) are divided into intervals (such as "high temperature" and "medium speed") according to business logic and converted into discrete variables. Missing value processing and outlier filtering are performed on the quality defect record data to construct a training data set: ;in, is the historical die-casting parameter vector, The quality defect type.

[0064] 2. Causal structure learning: (1) Scoring search method to build network skeleton: Use the Bayesian Information Criterion (BIC) scoring function to evaluate candidate network structures: .

[0065] Where G is the network structure, d is the number of parameters, n is the sample size, and D is the sample dataset. A greedy search algorithm is used to find the optimal structure, generating a directed acyclic graph (DAG). Nodes represent parameters / defects, and edges represent causal dependencies (e.g., "pouring temperature → pores").

[0066] (2) Conditional independence verification: Perform a chi-square test on the generated edges and calculate the statistic: .

[0067] like If the value exceeds the critical value, the edge is retained, indicating that there is a significant dependence between the parameter and the defect.

[0068] 3. Parameter learning and conditional probability estimation: Under a certain network structure G, the maximum likelihood estimation (MLE) is used to calculate the conditional probability table (CPT) of each node: For a discrete variable Y and its parent node set Pa(Y), each entry in the CPT is: .

[0069] For example, if "air pores" occur 30 times out of 100 samples with "pouring temperature = high temperature" and "filling speed = high speed", then: P(air pores = yes | temperature = high temperature, speed = high speed) = 0.3.

[0070] 4. Second association rule generation: Convert the conditional probability relationship in the Bayesian network into explicit rules: Direct causal rule: For the edge X→Y in the network, generate rule X→Y with the strength of conditional probability P(Y|X).

[0071] For example: the rule: "If the pouring temperature = high temperature, then the probability of porosity defect = 0.45" is recorded as: temperature = high temperature → porosity = yes, [strength = 0.45].

[0072] Multi-condition rule: For a node Y with multiple parent nodes, a joint condition rule is generated.

[0073] For example: the rule: "If holding time = short and mold temperature = low, then the probability of shrinkage defect = 0.8" is recorded as: (holding time = short ∧ mold temperature = low) → shrinkage = yes, [strength = 0.8].

[0074] 5. Quantification of the second association strength: Two strength indicators are assigned to each second association rule: (1) Conditional probability strength: directly use the conditional probability value P(Y|X) in CPT to reflect the direct impact of the parameter on the defect.

[0075] (2) Causal effect strength: The actual causal effect of the parameter change on the defect is quantified using the intervention probability P(Y|do(X)), which is calculated using the do-calculus: causal effect = P(Y|do(X=x1))-P(Y|do(X=x0)).

[0076] For example: when the pouring temperature is adjusted from low temperature to high temperature, the probability of air holes increases from 0.15 to 0.45, and the causal effect is 0.45-0.15=0.3.

[0077] 6. Rule screening and formation of the second association rule set: retain the second association rules with conditional probability strength ≥ 0.6 and causal effect strength ≥ 0.2 to ensure that the rules have actual predictive value. All verified second association rules are organized into the second association rule set.

[0078] An example of the second association rule set obtained through steps 1-6 above is shown in Table 3 below: Table 3:

[0079] It should be noted that the above Table 3 of the present invention is only used for illustrative purposes, and is not intended to limit the second association rule set in the present invention to only include the contents of the above Table 3.

[0080] Finally, the first and second association rule sets are merged, retaining the association rules supported by both algorithms as strong association rules. These rules are highly credible and form the core of the causal association comparison table. The specific processing process is as follows: 1. Unify the variables of the two types of rules into the same semantic space (for example, use the causal expression "die casting parameter → defect type").

[0081] For example: The first association rule: {high temperature, fast speed} → stomata (support = 0.7, confidence = 0.8).

[0082] Second association rule: high temperature → stomata (P(stomata|high temperature) = 0.6, causal effect = 0.3).

[0083] 2. Keep rules whose variable combinations and causal directions are completely consistent in both types of rules: Only when the antecedent→consequent direction of the first association rule completely matches the directed edge (e.g., A→B) of the second association rule is considered "jointly supported." Exclude any reverse associations (e.g., B→A) or undirected associations (e.g., A↔B) in the first association rule.

[0084] At the same time, for other first and second association rules that were not selected as strong association rules but whose association strengths exceeded the threshold, a different algorithm is used for reverse verification. These other first and second association rules that passed verification are also integrated into the strong association rules. A causal correlation comparison table is constructed based on the integrated strong association rules, ensuring that the rules in the table are comprehensive and accurate, effectively improving the reliability of quality predictions.

[0085] The Bayesian network algorithm was used to reverse-validate the other first association rules mentioned above. The specific process was as follows: 1. Rules that met the following conditions were extracted from the first association rule set: 1) Rules that were not retained by the fusion of the first and second association rules; 2) Rules whose first association strength exceeded a threshold: for example, the rule's confidence (Confidence) or lift (Lift) exceeded a preset threshold (e.g., confidence ≥ 0.7, lift ≥ 1.5). For example, the rule {Low pressure, low mold temperature} → shrinkage has a confidence of 0.8 (threshold 0.7), but there is no direct causal edge from low pressure → shrinkage or low mold temperature → shrinkage in the Bayesian network, so it was not included in the strong association rule.

[0086] 2. Convert the original quality defect record data into the structured data set required by the Bayesian network, including: all relevant variables (such as die-casting parameters: pressure, temperature; defect types: shrinkage, pores); variable values ​​(such as pressure "low / medium / high", defect "presence / absence"); independent and identically distributed observations between samples (each record corresponds to the parameters and defect results of a die-casting process).

[0087] 3. Bayesian network structure learning: exploring potential causal paths to initialize the network skeleton: using scoring search methods (such as greedy search, hill climbing algorithm) or constraint methods (such as PC algorithm) to learn the Bayesian network structure from the data and generate an undirected graph or directed acyclic graph (DAG).

[0088] Scoring functions (such as the BDeu score) are used to assess the fit of the structure to the data. A higher score indicates a more likely correct structure. For example, when exploring the relationship between low pressure and shrinkage porosity, an indirect causal path (including the mediating variable flow rate) was found: low pressure → slow metal flow rate → incomplete filling → shrinkage porosity.

[0089] Undirected edges are assigned directions using conditional independence tests or expert knowledge to ensure a loop-free network. If low pressure and shrinkage are conditionally independent at a given flow rate in the data, the path is low pressure → slow flow rate → shrinkage, not a direct relationship.

[0090] 4. Bayesian Network Parameter Learning: Quantifying probabilistic dependency parameter estimation: Using maximum likelihood estimation (MLE) or Bayesian estimation (e.g., Dirichlet priors) to calculate the conditional probability table (CPT) for each node in the network. For example, P(shrinkage|low pressure, slow flow) and P(shrinkage|low pressure, fast flow) are calculated to assess whether the direct effect of low pressure on shrinkage is mediated by flow rate.

[0091] For the variable pairs in the target rule (e.g., low pressure → shrinkage), calculate: (1) conditional probability: P(consequent|antecedent), reflecting the probability of the consequent when the antecedent occurs (e.g., P(shrinkage|low pressure)); (2) intervention probability: P(consequent|do(antecedent)) is calculated through causal inference (e.g., backdoor adjustment, frontdoor adjustment), eliminating the influence of confounding variables and quantifying the direct causal effect of the antecedent on the consequent.

[0092] 5. Rule Verification: Double screening based on causal logic and probability threshold Verification condition setting: Other first association rules that meet the following conditions are considered to have “passed verification”: (1) There is a causal path: There is a direct or indirect causal path from the antecedent to the consequent of the rule in the Bayesian network (e.g., antecedent → mediating variable → consequent).

[0093] (2) The strength of the probability association meets the standard: the conditional probability P(consequent|antecedent) ≥ threshold 1 (e.g., 0.6); the causal effect after the intervention (e.g., relative risk RR or attributable risk AR) ≥ threshold 2 (e.g., RR ≥ 1.8).

[0094] For example, the first association strength for the rule {low pressure} → shrinkage is above the threshold, but there is no direct edge in the Bayesian network: low pressure → shrinkage. However, an indirect path exists: low pressure → slow flow rate → shrinkage. The calculated result is: P(shrinkage|low pressure) = 0.75 (threshold 0.6). After intervention (P(shrinkage|do(low pressure)) = 0.72) (accounting for flow rate as a confounding variable), the causal effect RR = 1.9 (threshold 1.8). Conclusion: The rule passes validation because an indirect causal path exists and the probability strength meets the threshold.

[0095] Filter invalid rules: If the antecedent and consequent of a rule are conditionally independent in the Bayesian network (i.e., there is no causal path and the correlation is caused by confounding variables), then reject the other first association rule.

[0096] For example, the confidence level of the rule {Low mold temperature} → Porosity is 0.75. However, the correlation between low mold temperature and porosity in the network is driven by the common cause, high gas content in the alloy liquid. There is no causal relationship between the two, so the rule fails verification.

[0097] 6. Output verification results: Generate an interpretable subset of first association rules. Verified rule set: Integrate the verified rules, and each first association rule is accompanied by: (1) a description of the causal path (for example, "low pressure indirectly causes shrinkage by reducing the flow rate of molten metal"); (2) probabilistic evidence: conditional probability, causal effect value; (3) an explanation of the difference with the strong association rule (for example, "it is an indirect causal relationship and needs to be regulated by a mediating variable").

[0098] The Apriori algorithm was used to reverse-validate the other secondary association rules mentioned above. The specific process is as follows: 1. Rules that meet specific conditions are selected from the secondary association rule set. These rules must meet the following requirements: first, they must not be retained during the association rule fusion process; second, their secondary association strength (e.g., confidence level, lift, etc.) must be above a pre-set threshold. For example, the rule {Fast cooling rate, low release agent dosage} → surface defects has a confidence level of 0.8 (threshold set at 0.7). However, in the existing strong association rule system, there are no directly recognized causal relationships such as "Fast cooling rate → surface defects" or "Low release agent dosage → surface defects." Therefore, this secondary association rule meets the screening criteria for this reverse validation.

[0099] 2. Convert the original quality defect record data into a structured dataset suitable for processing by the Apriori algorithm. This structured dataset includes all relevant variables, such as die-casting process parameters (such as pressure, temperature, cooling rate, and release agent dosage) and defect types (such as surface defects, pores, and shrinkage). It also clearly defines the value of each variable, such as "low, medium, or high" for pressure and "presence or absence" for defects. The data is ensured to consist of independent and identically distributed sample observations, meaning that each record corresponds to the actual values ​​of each parameter in a die-casting process and the final defect result.

[0100] 3. Apriori algorithm forward association rule mining: Generate candidate item sets and frequent item sets: First, generate candidate item sets from the data set. Start with a single item, count the frequency of each item in the data set, that is, calculate the support (support = number of transactions containing the item set / total number of transactions), and filter out items with support greater than or equal to the support threshold (set as s threshold) items to form frequent 1-item sets. Then, based on the frequent 1-item sets, the items in them are combined to generate candidate 2-item sets. The support of these candidate 2-item sets is calculated again, and the items with support greater than or equal to s are retained. threshold as a frequent 2-itemset.

[0101] In the above way, higher-order candidate item sets are continuously generated and frequent item sets are filtered out until no new frequent item sets that meet the support threshold can be generated.

[0102] Generate positive association rules and calculate confidence: Generate all possible positive association rules from the frequent item sets. For example, for the frequent item set {fast cooling rate, low release agent dosage, surface defects}, you can generate a positive rule like {fast cooling rate, low release agent dosage} → {surface defects}. Then calculate the confidence of each positive association rule (confidence = support (antecedent ∪ consequent) / support (antecedent)).

[0103] 4. Reverse verification process: Verification of the inference of the antecedent from the consequent (Y↛X): Generate reverse candidate item sets: Based on the transaction data containing items related to the consequent (e.g., "surface defects"), generate candidate item sets containing items related to the consequent according to the method of generating candidate item sets above, such as {surface defects, fast cooling speed}, etc. Calculate the support of these candidate item sets and select those with support greater than or equal to s. threshold Frequent itemsets.

[0104] Construct reverse association rules and calculate confidence: Generate reverse association rules from these frequent item sets, such as {surface defect} → {fast cooling rate}. Calculate the confidence of each reverse association rule (using the same formula as for forward rule confidence calculation).

[0105] Judging the rationality of the rules: If the confidence of all these reverse association rules is less than the confidence threshold c threshold , then it means that the situation of inferring the antecedent from the posterior is not established, which meets the requirements of reverse verification; if there is a confidence level greater than or equal to c threshold If the reverse rule is found, it indicates that the original association rule may be logically inaccurate and needs further in-depth analysis.

[0106] Verification of the non-predecessor-postponed non-consequent (¬X ↛ ¬Y): Generate candidate itemsets for the non-predecessor: Find transactions in the dataset that do not contain items related to the predecessor (e.g., "fast cooling speed, low release agent dosage"). Based on these transactions, generate candidate itemsets, such as {normal cooling speed, normal release agent dosage}. Calculate the support of the candidate itemsets and determine the frequent itemsets.

[0107] Construct reverse association rules and calculate confidence levels: Generate association rules such as {normal cooling rate, normal release agent dosage} → {no surface defects}, which are rules that infer non-preceding conditions from non-consequential conditions. Calculate their confidence levels.

[0108] Judge the rationality of the rules: if the confidence of these reverse association rules is less than the confidence threshold c threshold , indicating that non-preconditions cannot lead to non-consequents, which is consistent with the expectation of reverse verification; if the confidence of a rule is greater than or equal to c threshold , then it is necessary to review and adjust the original association rules.

[0109] 5. Based on the results of the two aforementioned reverse validations, conduct a comprehensive evaluation of each secondary association rule. If both Y ↛ X and ¬X ↛ ¬Y meet expectations (i.e., the reverse rule confidence is less than the threshold), the association rule can be considered reasonable and reliable in terms of reverse logic. If any of these conditions are not met, the original rule needs to be revised or marked as requiring further research. The validation results are collated and output, and then integrated into the strong association rules to generate a causal correlation table.

[0110] Through the above process, the present invention realizes reverse verification of other first association rules and second association rules that meet the above conditions, supplements the "potential causal rules" to the end of the causal association comparison table, and helps to improve the accuracy of subsequent quality defect prediction results.

[0111] As an example, the quality defect prediction model corresponding to the defect prediction type is used to perform prediction processing on the prediction basic data to obtain the defect probability of the aluminum alloy die-casting corresponding to the defect prediction type, including: obtaining a multimodal input sequence based on the prediction basic data; processing the multimodal input sequence through a hybrid attention mechanism to obtain a local feature matrix and a global feature matrix; calculating the association matrix between the local feature matrix and the global feature matrix, generating a fusion feature matrix based on the association matrix, and generating an enhanced feature vector based on the fusion feature matrix and temporal context information; the temporal context information is extracted through a self-attention mechanism; mapping the enhanced feature vector to the defect probability space to obtain the defect probability of the aluminum alloy die-casting corresponding to the defect prediction type.

[0112] The quality defect prediction model of the present invention adopts the Transformer-Enhanced hybrid attention network model, including a local attention module, a global attention module and a multi-layer perceptron, such as Figure 2 shown.

[0113] First, the temporal parameters and static parameters in the input prediction basic data are encoded as temporal feature vectors and static feature vectors, respectively. These feature vectors are then mapped to a unified feature space to form a multimodal input sequence. Specifically: Temporal parameter encoding: For time-varying parameters such as temperature and pressure, 1D convolution is used to extract local feature patterns (e.g., temperature drops, pressure fluctuations), and positional encoding is then superimposed to preserve temporal sequence information. Static parameter encoding: For fixed parameters such as alloy composition and mold structure, an embedding layer is used to map them into low-dimensional semantic vectors (e.g., mapping the "AlSi9Cu3" alloy to a specific vector representation).

[0114] Project the time series feature vector and the static feature vector to the same dimension (e.g. 256 dimensions) through linear transformation to form a multimodal input sequence X∈R T×d .

[0115] Then, the multimodal input sequence is processed by the hybrid attention mechanism to obtain the local feature matrix and the global feature matrix. Specifically: (1) The local attention module transforms the multimodal input sequence into a query matrix, a key matrix and a value matrix through linear transformation, calculates the correlation strength between time steps through the scaled dot product attention formula, forms a temporal correlation weight matrix, and uses the weight matrix to perform weighted fusion on the value matrix to generate a local feature matrix. Specifically: The multimodal input sequence X is transformed into a query matrix, a key matrix and a value matrix through the weight matrix W. Q 、W K 、W V Converted into query matrix Q, key matrix K and value matrix V respectively.

[0116] The dependencies between time steps are calculated using the scaled dot product formula: .

[0117] in, Indicates time The parameters of the moment The influence strength of the parameters; is the query vector at time step Vector representation of moments; yes The transpose of is in the time step Vector representation of moments; is the dimension of the key vector.

[0118] Generate local feature matrix by weighted summation ,capturing the timing-dependent patterns in the die-casting process (e.g., the effect of temperature changes on subsequent pressure).

[0119] (2) The global attention module converts the causal association table into a directed graph, embeds the nodes into a low-dimensional space, and obtains node representations. For each parameter node and defect node, the graph attention mechanism calculates the causal association weight. Based on the causal association weight, the node representation is updated by aggregating the information of neighboring nodes to generate a global feature matrix. Specifically, the causal association table (e.g., "Insufficient holding time → shrinkage") is converted into a directed graph G, where nodes represent parameters or defects and edge weights represent the strength of the association (e.g., conditional probability).

[0120] For each parameter node i and defect node j, the causal association weight is calculated using the following formula: .

[0121] in, is the edge weight between nodes i and j in the knowledge graph, is the knowledge fusion coefficient; is a learnable weight vector; is a learnable weight matrix, 、 Represent the feature vectors of node i and node j respectively; Represents the set of neighbor nodes of node i.

[0122] Aggregate neighbor information based on causal association weights to generate a global feature matrix .

[0123] Then, the correlation matrix between the local feature matrix and the global feature matrix is ​​calculated, and the global causal knowledge is integrated into the local feature matrix through the correlation matrix to generate a fused feature matrix; the fused feature matrix is ​​globally averaged pooled in the time dimension, and the temporal context information is extracted through the self-attention mechanism to generate an enhanced feature vector. Specifically: the correlation matrix between the local feature and the global feature is calculated through the attention mechanism ,in, Represents the association strength between the parameter at time t and node i in the knowledge graph.

[0124] Incorporating global causal knowledge into local features: ;in, Represents the fusion feature matrix At time step The eigenvector when ; Represents the eigenvector of the local feature matrix at time step t; Represents the element value of the cross-modal attention association matrix corresponding to the knowledge graph node i at time step t; Represents the eigenvector corresponding to node i in the global feature matrix.

[0125] For example, if , the causal knowledge of “holding pressure” is strongly injected into the local features at time t.

[0126] The fused feature matrix Averaged over the time dimension to obtain a fixed-length vector .

[0127] Further extract temporal context information through self-attention: ,Through this operation, the model can focus on the time segments that are most critical for defect prediction (such as parameter changes in the filling stage).

[0128] Finally, the enhanced feature vector is mapped to the defect probability space through a multi-layer perceptron, and the mapped probability is normalized to obtain the defect probability of the aluminum alloy die casting corresponding to the defect prediction type. Mapping to defect probability space: ;in, , is the number of defect types (e.g. pores, shrinkage, cracks); is the learnable weight coefficient; is the bias coefficient.

[0129] The linear output is converted into a probability distribution using, for example, a Softmax function, that is, the defect probability of the aluminum alloy die casting corresponding to the defect prediction type is obtained.

[0130] As an example, the multimodal input sequence is processed through a hybrid attention mechanism to obtain a local feature matrix, including: the local attention module generates a query matrix, a key matrix and a value matrix through a linear transformation of the multimodal input sequence, and calculates the correlation strength between time steps through the scaled dot product attention formula; calculates the correlation strength trend represented by all historical time steps, generates a corresponding optimization coefficient according to the correlation strength trend, and uses the optimization coefficient to correct each of the correlation strengths to form a time series correlation weight matrix.

[0131] During the die-casting process, the correlation strength of timing parameters may exhibit a trend over time (for example, due to equipment aging or abnormal intermittent fluctuations in equipment, the correlation between parameters gradually increases or decreases). However, the standard self-attention mechanism treats all historical time steps equally and cannot capture the dynamic trend of correlation strength (for example, the influence of early time steps may decay with production). Therefore, the present invention analyzes the changing patterns of historical correlation strength and dynamically adjusts the attention weights to improve the model's adaptability to timing trends.

[0132] First, the local attention module calculates the association strength between time steps in a conventional way, that is, the multimodal input sequence is linearly transformed to generate the query matrix, key matrix and value matrix, and the association strength between time steps is calculated by the scaled dot product attention formula.

[0133] Then, a trend analysis is performed on the correlation strength corresponding to all historical time steps (using methods such as sliding average and polynomial fitting) to obtain the linear increasing / decreasing trend, exponential decay trend, etc. of the correlation strength, and the optimization coefficient is generated according to the corresponding trend.

[0134] For example, if the correlation strength of the first 30 time steps is found to be exponentially decaying, it can be fitted as ,in, is the attenuation coefficient, is the historical time step.

[0135] According to the trend analysis results, for each historical time step Generate optimization coefficients (e.g., a trend enhancement or decay coefficient). For example, for early time steps (e.g. Give smaller , suppressing its correlation strength; giving a larger , enhancing its impact.

[0136] All association intensities are modified using the optimization coefficients to form a time series association weight matrix.

[0137] This embodiment generates optimization coefficients through trend analysis to correct the association weights between time steps, solving the defect that standard self-attention cannot capture the dynamic trend of temporal association strength. It can dynamically adapt to the temporal changes in the production process and improve prediction accuracy and interpretability.

[0138] like Figure 3 As shown, an embodiment of the present invention further provides an aluminum alloy die casting quality prediction system 10 , which includes a causal association combing module 100 , a key parameter determination module 200 , a die casting parameter acquisition module 300 , and a quality defect prediction module 400 .

[0139] The causal correlation combing module 100 collects quality defect record data of aluminum alloy die-casting parts, performs causal correlation analysis on each quality defect record data, and obtains a causal correlation comparison table; the quality defect record data includes quality defect types and corresponding historical die-casting parameters.

[0140] The key parameter determination module 200 determines each defect prediction type of this round, compares each defect prediction type with the causal association comparison table, and obtains several key parameter types corresponding to any defect prediction type.

[0141] The die-casting parameter acquisition module 300 acquires real-time die-casting parameters during the aluminum alloy die-casting process, and selects data corresponding to each key parameter type to form prediction basic data.

[0142] The quality defect prediction module 400 uses the quality defect prediction model corresponding to the defect prediction type to perform prediction processing on the prediction basic data, obtains the defect probability of the aluminum alloy die casting corresponding to the defect prediction type, and outputs an alarm signal when the defect probability is higher than a probability threshold.

[0143] As an example, the key parameter determination module 200 specifically: retrieves several trial production data corresponding to this round, and the quality inspection results corresponding to each trial production data, and determines each defect type in the quality inspection results corresponding to the last trial production as the defect prediction type.

[0144] As an example, the causal association combing module 100 specifically: uses the Apriori algorithm to perform preliminary association rule mining on each of the quality defect record data to generate a first association rule set; the first association strength corresponding to each first association rule association in the first association rule set; uses the Bayesian network algorithm to perform causal structure learning on each of the quality defect record data, constructs a probabilistic dependency relationship between die-casting parameters and defect types, and generates a second association rule set; the second association strength corresponding to each second association rule association in the second association rule set; cross-validates and fuses the first association rule set and the second association rule set to retain strong association rules supported by both algorithms; uses the Bayesian network algorithm to verify the first association rules in the first association rule set that do not belong to the strong association rules and whose first association strength is higher than the threshold, and uses the Apriori algorithm to verify the second association rules in the second association rule set that do not belong to the strong association rules and whose second association strength is higher than the threshold, to obtain several first association rules and second association rules that meet the verification conditions, and combines them with the strong association rules to construct the causal association comparison table.

[0145] As an example, the quality defect prediction module 400 specifically: obtains a multimodal input sequence based on the prediction basic data; processes the multimodal input sequence through a hybrid attention mechanism to obtain a local feature matrix and a global feature matrix; calculates the correlation matrix between the local feature matrix and the global feature matrix, generates a fusion feature matrix based on the correlation matrix, and generates an enhanced feature vector based on the fusion feature matrix and temporal context information; the temporal context information is extracted through a self-attention mechanism; and maps the enhanced feature vector to a defect probability space to obtain the defect probability of the aluminum alloy die-casting corresponding to the defect prediction type.

[0146] As an example, the quality defect prediction module 400, specifically: the local attention module generates a query matrix, a key matrix and a value matrix through linear transformation of the multimodal input sequence, and calculates the correlation strength between time steps through the scaled dot product attention formula; calculates the correlation strength trend represented by all historical time steps, generates a corresponding optimization coefficient according to the correlation strength trend, and uses the optimization coefficient to correct each of the correlation strengths to form a time series correlation weight matrix.

[0147] An embodiment of the present invention further provides an electronic device comprising: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program implements any of the aforementioned methods when executed by the processor.

[0148] An embodiment of the present invention further provides a computer storage medium storing a computer program that can be executed by a processor to implement any of the methods described above.

[0149] An embodiment of the present invention further provides a computer program product, which includes a computer program that can be executed by a processor to implement any of the methods described above.

[0150] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the composition and steps of each example according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0151] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for predicting the quality of aluminum alloy die castings, characterized by: The method includes the following steps: collecting quality defect record data of aluminum alloy die-casting parts, performing causal correlation analysis on each quality defect record data, and obtaining a causal correlation comparison table; the quality defect record data includes quality defect types and corresponding historical die-casting parameters; determining each defect prediction type of this round, comparing each defect prediction type with the causal correlation comparison table, and obtaining several key parameter types corresponding to any defect prediction type; collecting real-time die-casting parameters in the aluminum alloy die-casting process, and selecting data corresponding to each of the key parameter types to form prediction basic data; using the quality defect prediction model corresponding to the defect prediction type to perform prediction processing on the prediction basic data, and obtain the defect probability of the aluminum alloy die-casting parts corresponding to the defect prediction type, and outputting an alarm signal when the defect probability is higher than the probability threshold.

2. The method for predicting quality of aluminum alloy die castings according to claim 1, wherein: Determining each defect prediction type of this round includes: retrieving a number of trial production data corresponding to this round, and quality inspection results corresponding to each trial production data, and determining each defect type in the quality inspection result corresponding to the last trial production as the defect prediction type.

3. The method for predicting quality of aluminum alloy die castings according to claim 1, wherein: A causal association analysis is performed on each quality defect record data to obtain a causal association comparison table, including: using the Apriori algorithm to perform preliminary association rule mining on each quality defect record data to generate a first association rule set; each first association rule in the first association rule set has a corresponding first association strength; using the Bayesian network algorithm to perform causal structure learning on each quality defect record data, construct a probabilistic dependency relationship between die-casting parameters and defect types, and generate a second association rule set; each second association rule in the second association rule set has a corresponding second association strength; cross-validation and fusion are performed on the first association rule set and the second association rule set to retain strong association rules supported by both algorithms; using the Bayesian network algorithm to verify the first association rules in the first association rule set that are not strong association rules and whose first association strength is higher than a threshold, and using the Apriori algorithm to verify the second association rules in the second association rule set that are not strong association rules and whose second association strength is higher than a threshold, to obtain several first association rules and second association rules that meet the verification conditions, and combine them with the strong association rules to construct the causal association comparison table.

4. The method for predicting the quality of aluminum alloy die castings according to claim 3, wherein: The quality defect prediction model corresponding to the defect prediction type is used to perform prediction processing on the prediction basic data to obtain the defect probability of the aluminum alloy die-casting corresponding to the defect prediction type, including: obtaining a multimodal input sequence according to the prediction basic data; processing the multimodal input sequence through a hybrid attention mechanism to obtain a local feature matrix and a global feature matrix; calculating the association matrix between the local feature matrix and the global feature matrix, generating a fused feature matrix based on the association matrix, and generating an enhanced feature vector based on the fused feature matrix and temporal context information; the temporal context information is extracted through a self-attention mechanism; mapping the enhanced feature vector to the defect probability space to obtain the defect probability of the aluminum alloy die-casting corresponding to the defect prediction type.

5. The method for predicting quality of aluminum alloy die castings according to claim 4, wherein: The multimodal input sequence is processed through a hybrid attention mechanism to obtain a local feature matrix, including: a local attention module generates a query matrix, a key matrix and a value matrix through linear transformation of the multimodal input sequence, and calculates the association strength between time steps through a scaled dot product attention formula; calculates the association strength trend represented by all historical time steps, generates a corresponding optimization coefficient according to the association strength trend, and uses the optimization coefficient to correct each of the association strengths to form a temporal association weight matrix.

6. A quality prediction system for aluminum alloy die castings, characterized in that: The system includes a causal correlation combing module, a key parameter determination module, a die-casting parameter collection module, and a quality defect prediction module; the causal correlation combing module collects quality defect record data of aluminum alloy die-casting parts, performs causal correlation analysis on each quality defect record data, and obtains a causal correlation comparison table; the quality defect record data includes quality defect types and corresponding historical die-casting parameters; the key parameter determination module determines each defect prediction type in this round, compares each defect prediction type with the causal correlation comparison table, and obtains several key parameter types corresponding to any defect prediction type; The die-casting parameter acquisition module collects real-time die-casting parameters during the aluminum alloy die-casting process, and selects data corresponding to each key parameter type to form prediction basic data; the quality defect prediction module uses the quality defect prediction model corresponding to the defect prediction type to perform prediction processing on the prediction basic data, and obtains the defect probability of the aluminum alloy die-casting corresponding to the defect prediction type, and outputs an alarm signal when the defect probability is higher than the probability threshold.

7. The aluminum alloy die casting quality prediction system according to claim 6, characterized in that: The key parameter determination module specifically retrieves several trial production data corresponding to this round, and the quality inspection results corresponding to each trial production data, and determines each defect type in the quality inspection result corresponding to the last trial production as the defect prediction type.

8. An electronic device, characterized in that: The electronic device includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program implements the method according to any one of claims 1 to 5 when executed by the processor.

9. A computer storage medium, characterized in that: The computer storage medium stores a computer program that can be executed by a processor to implement the method according to any one of claims 1 to 5.

10. A computer program product, characterized in that: The computer program product comprises a computer program that can be executed by a processor to implement the method according to any one of claims 1 to 5.

Images