A quality traceability method and device for an air conditioner top integrated machine production line

Abstract

The application discloses a quality traceability method and device for an air conditioner top-mounted integrated machine production line, aiming at the problems of multi-source heterogeneous production data time sequence fusion and abnormal cause and effect traceability, and extracts behavior mode signatures containing operation sequence topology, state jump frequency and abnormal context association through unified time sequence alignment and window slicing of structured sensors, equipment logs and operation records. Further, a standardized process semantic fingerprint is constructed by using a hash vector compression, fast similarity comparison and efficient storage are realized. Based on historical good product data, combining information theory and statistical cause and effect test, a dynamic cause and effect atlas between processes is established, abnormal quality signals are responded in real time, and an explainable attribution path is output through limited depth traversal and semantic deviation determination. The scheme improves the normalization modeling precision of heterogeneous data and strengthens the cause and effect tracing explainability of quality abnormalities.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and production line process quality control technology, and in particular to a quality traceability method and device for an air conditioner rooftop integrated unit production line. Background Technology

[0002] Currently, in the manufacturing industry, especially in the production of complex equipment such as air conditioner ceiling-mounted units, full-process monitoring and traceability analysis of product quality has become a core requirement for improving process level, ensuring finished product consistency, and closed-loop quality management.

[0003] The industry still relies on a rudimentary approach of traditional data statistics combined with manual inspection for quality traceability of such products, lacking an intelligent traceability system. Existing solutions only rely on sensors to collect equipment operating parameters, combined with manual operation, inspection, and quality control records to complete simple data archiving. When quality anomalies occur, technical personnel rely entirely on manually comparing parameter thresholds, reviewing production records, and combining experience to determine the cause. This approach can only address explicit and recurring quality issues such as parameter exceedances and equipment malfunctions, and its applicability to various scenarios is extremely limited.

[0004] Existing technologies suffer from several key shortcomings. First, production condition data, operational data, and quality inspection data are stored in fragmented, multi-source formats, lacking standardized integration and time alignment mechanisms. Anomaly investigation requires manual data integration and comparison, resulting in a large workload, slow response, and extremely low efficiency in closing the quality loop. Second, under the dynamic production characteristics of small batches and multiple production runs, process and equipment parameters are frequently adjusted. Traditional manual traceability rules cannot adapt to changes in operating conditions, exhibiting poor adaptability and robustness, making it difficult to guarantee traceability accuracy. Third, existing methods rely solely on numerical thresholds to determine anomalies, ignoring implicit influencing factors such as operational behavior, process state transitions, and multi-parameter coupling. This fails to uncover deep semantic features of the production process, making it difficult to locate hidden quality causes, leading to frequent missed and false diagnoses. Furthermore, traditional full-domain data backtracking and investigation models are resource-intensive and slow in traceability, easily delaying production cycles and causing batches of defective products, making them unsuitable for the needs of refined intelligent manufacturing control. Summary of the Invention

[0005] This application provides a quality traceability method and apparatus for an air conditioner rooftop unit production line, aiming to solve one of the problems or issues of the prior art mentioned in the background.

[0006] This application provides a quality traceability method for an air conditioner ceiling-mounted integrated unit production line, specifically including: Acquire multi-source heterogeneous data from each process unit of the air conditioner rooftop unit production line, perform unified semantic anchoring processing on the multi-source heterogeneous data, and generate an original behavior pattern signature set. Based on the original behavioral pattern signature set, a hash vector representation of each process unit is generated, a process semantic fingerprint database is constructed, and the original behavioral pattern signature set is transformed into a process semantic fingerprint sequence. The cross-process influence relationship is statistically verified using the process semantic fingerprint sequence. The discrete process semantic fingerprint sequences are mapped to generate a basic causal skeleton graph structure where nodes are process semantic fingerprints and edges are lag correlations. In response to the quality anomaly signal detected in real time, the fingerprint of the current abnormal process is used as the query starting point. A restricted depth-first traversal is performed on the basic causal skeleton graph and adjacent candidate process semantic fingerprints are dynamically loaded. The semantic deviation between the current abnormal process fingerprint and the candidate process semantic fingerprint is calculated, and a set of dynamically activated edges is generated. Based on the discrete behavioral features shared in the set of dynamically activated edges, local causal strength re-evaluation is performed and a dynamic causal graph with short-term memory factors is generated, which is then transformed into a dynamic causal graph link. Based on the short-term memory factor threshold judgment result of the dynamic causal graph link, the occasional interference edge connection below the set threshold is weakened and the process semantic chain composed of process semantic fingerprint nodes and dynamic confidence scores is output, and the dynamic causal graph link is parsed into the final quality problem attribution path.

[0007] This application provides a quality traceability device for an air conditioner rooftop unit production line, specifically comprising: The barcode scanner is installed in each process unit of the production line to generate a unique traceability code for each air conditioner top-mounted unit. The industrial control terminal is connected to the barcode scanner and is used to execute the above-mentioned quality traceability method for the air conditioner top-mounted integrated unit production line, collect the process parameters generated by each process unit, and associate the process parameters generated by each process unit with the traceability code in real time. The traceability device is connected to the industrial control terminal and is used to retrieve, query, and display the status of each process.

[0008] The beneficial effects of the quality traceability method and system for an integrated rooftop air conditioner production line provided in this application are as follows: By standardizing and integrating multi-source heterogeneous data from various process units using unified semantic anchoring, this invention solves the problems of traditional data fragmentation and the need for manual integration and comparison. Relying on process behavior pattern signatures, semantic fingerprint databases, and dynamic causal graph evolution mechanisms, it abandons the traditional full-domain data backtracking and static threshold attribution modes. This significantly reduces data processing resource overhead, improves the response speed of anomaly tracing and the efficiency of quality problem closure, and can adapt to complex production conditions such as small batches, multiple batches, and dynamic adjustment of process parameters on production lines, exhibiting excellent adaptability and traceability stability. At the same time, this invention breaks through the limitations of traditional methods that rely solely on numerical thresholds to determine anomalies. Through semantic deviation calculation and discrete behavioral feature analysis, it deeply mines hidden influencing factors such as operational behavior, process state transitions, and multi-parameter coupling, achieving fine-grained, interpretable, and accurate positioning of quality anomalies. This effectively reduces missed and false judgments, enabling rapid and accurate investigation of production anomalies, avoiding the production of batches of defective products, ensuring the stable and efficient operation of the production line, and significantly improving the intelligent quality control level and overall production efficiency of complex equipment production lines. Attached Figure Description

[0009] Figure 1 This is a main flowchart of a quality traceability method for an air conditioner rooftop unit production line.

[0010] Figure 2 This is a sub-flowchart of a quality traceability method for an air conditioner rooftop unit production line.

[0011] Figure 3 This is another sub-flowchart of a quality traceability method for an air conditioner rooftop unit production line. Detailed Implementation

[0012] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0013] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0014] like Figure 1 As shown, this application provides a quality traceability method for an air conditioner ceiling-mounted integrated unit production line, specifically including: S1: Obtain multi-source heterogeneous data from each process unit of the air conditioner rooftop integrated unit production line, perform unified semantic anchoring processing on the multi-source heterogeneous data, and generate an original behavior pattern signature set. S2: Generate a hash vector representation for each process unit based on the original behavior pattern signature set, construct a process semantic fingerprint database, and transform the original behavior pattern signature set into a process semantic fingerprint sequence; S3: Utilize the process semantic fingerprint sequence to statistically verify the cross-process influence relationship, and map the discrete process semantic fingerprint sequence into a basic causal skeleton graph structure with nodes as process semantic fingerprints and edges as lag correlations; S4: In response to the quality anomaly signal detected in real time, take the fingerprint of the current abnormal process as the query starting point, perform a restricted depth-first traversal on the basic causal skeleton graph and dynamically load the semantic fingerprints of adjacent candidate processes, calculate the semantic deviation between the current abnormal process fingerprint and the semantic fingerprints of candidate processes, and generate a set of dynamically activated edges. S5: Based on the discrete behavioral features shared in the set of dynamically activated edges, perform local causal strength re-evaluation and generate a dynamic causal graph with short-term memory factors, which is then transformed into a dynamic causal graph link. S6: Based on the short-term memory factor threshold judgment result of the dynamic causal graph link, weaken the occasional interference edge connection below the set threshold and output the process semantic chain composed of process semantic fingerprint nodes and dynamic confidence scores, and parse the dynamic causal graph link into the final quality problem attribution path.

[0015] S1: Acquire multi-source heterogeneous data from each process unit of the air conditioner rooftop integrated unit production line, perform unified semantic anchoring processing on the multi-source heterogeneous data, and generate an original behavioral pattern signature set. Specifically, this includes: S1.1: Perform multi-source concurrent acquisition and timestamp alignment processing on the structured sensor data streams, semi-structured equipment log messages, and unstructured operation record sequences deployed at key workstations of the air conditioner rooftop integrated unit production line to generate a full original dataset containing torque timing curves, PLC status code sequences, and manual scanning action time points.

[0016] Lightweight data acquisition agents are deployed at the edge of each key workstation on the production line, connecting to the analog signal channel of the torque sensor, the Ethernet port of the PLC controller, and the wireless interface of the handheld barcode scanner, respectively, and establishing independent data buffer queues to isolate the transmission delay differences of different communication protocols.

[0017] High-frequency sampling and capture are performed on the structured sensor data stream. The sampling frequency is set to 1kHz to obtain the torque value sequence on a continuous time axis, and microsecond-level hardware clock stamps are added to establish an absolute time reference.

[0018] The semi-structured device log messages are parsed according to the protocol, the status word and alarm ID field of the PLC internal register are extracted, discrete status transition events are mapped into standardized logical status codes, and the precise system time of the status change is recorded.

[0019] Event-driven capture is performed on unstructured operation record sequences. The barcode scanner trigger signal is monitored, and the occurrence time of each scanning action and the associated work order identifier are recorded to form a sparse but critical human intervention timestamp sequence.

[0020] The global network time protocol NTP is used to periodically synchronize and calibrate the system clocks of each edge node, eliminating clock drift errors in the distributed acquisition environment and ensuring that all data sources share a unified time reference system.

[0021] Based on a unified time reference system, and using the high-frequency time axis of structured sensor data as the main framework, the low-frequency PLC status code sequence is mapped to the corresponding time scale through linear interpolation, thereby achieving time alignment between status data and analog data.

[0022] By inserting the time points of manual barcode scanning actions as discrete event markers into the aligned timeline, a three-dimensional heterogeneous data index structure containing continuous physical quantities, discrete logic states, and manual operation events is constructed.

[0023] By using the above-mentioned multi-source concurrent acquisition and clock synchronization alignment processing method, the dispersed and heterogeneous raw signals are transformed into a full raw dataset with strict consistency in time dimension, providing a high-precision time-series basis for subsequent data slicing based on standard beat windows, and significantly improving the spatiotemporal consistency of multimodal data fusion.

[0024] S1.2: Based on the production process cycle parameters of the full original dataset, perform a sliding window cutting operation on the continuous time axis to generate a set of independent process data slices that strictly correspond to the standard production cycle, ensuring that each data slice completely covers the entire operation process of a single process unit in the time dimension.

[0025] Receive the full original dataset aligned with timestamps, and extract the preset standard production cycle parameters for each key workstation as the basis for the step size and width of the sliding window.

[0026] Based on the process division logic of the air conditioner rooftop unit assembly process, the start trigger signal and termination confirmation signal corresponding to each independent process unit are determined, and a process boundary judgment rule set is constructed.

[0027] Using the time point of the initial trigger signal as the anchor point, the initial candidate data slice interval is generated by truncating a duration equal to the standard beat period along the continuous time axis.

[0028] Integrity verification is performed on multi-source heterogeneous data within the initial candidate data slice interval to detect whether there are data gaps caused by sensor packet loss or communication delay.

[0029] If missing data is detected, a linear interpolation algorithm is used to complete the structured sensor data, and the slice is marked as a low-confidence state awaiting verification.

[0030] Based on the process boundary determination rule set, verify whether the termination confirmation signal falls within the current slice interval to ensure closed-loop coverage of the physical operation process.

[0031] Calculate the percentage of steady-state duration of key process parameters within the slice, and encapsulate the verified and completed data slices into independent process data slice objects, attaching process ID, timestamp range, and confidence label.

[0032] Reorganize all independent process data slice objects according to the production batch order to generate a set of independent process data slices that strictly correspond to the standard production cycle.

[0033] By using the sliding window cutting and integrity verification methods described above, the continuous time-series data from the previous step is transformed into a set of discretized and standardized independent process data slices, achieving semantic alignment of multi-source heterogeneous data in the process dimension and providing a unified time benchmark for subsequent feature extraction.

[0034] For example, for the condenser assembly station, the standard cycle time is set to 45 seconds. The system uses the "clamping" signal from the PLC as the starting anchor point and captures torque, current, and PLC status code data within the following 45 seconds. If torque data is missing between the 12th and 13th seconds, linear interpolation is performed using adjacent sampling points to complete the data. The cumulative duration of torque fluctuations less than 0.5 Nm within this 45-second period is calculated to be 38 seconds, resulting in a steady-state percentage R of 0.84. Finally, an independent slice containing complete time-series data is generated to ensure that subsequent causal analysis is based on a complete single-process operation cycle.

[0035] S1.3: Using predefined industrial protocol parsing rules, the semi-structured equipment log messages in the independent process data slice set are subjected to field extraction and state mapping to generate discrete state event sequences that characterize the logical transitions of equipment operation, and non-standardized alarm IDs and error codes are transformed into computable logical state vectors.

[0036] For semi-structured equipment log messages in the independent process data slice set, field extraction operations are performed according to predefined industrial communication protocol parsing rules. For common industrial protocols such as Modbus TCP or OPC UA, the specific register address or object identifier carrying equipment status information in the message is located, irrelevant protocol headers and check bits are stripped off, and the original status code field is extracted.

[0037] The extracted raw status code fields are standardized and mapped to establish a mapping dictionary between non-standard alarm IDs and the internal unified logical state space. Error codes specific to devices from different manufacturers (such as E0x1A and Err_204) are converted into unified discrete state enumeration values, eliminating the semantic gap between heterogeneous devices and generating standardized discrete state event sequences.

[0038] Based on standardized discrete state event sequences, key transition points in the device's operational logic are identified. Changes in state values ​​between adjacent timestamps are detected, and the instant in which the state changes is marked as a logical transition event. The source state before the transition and the target state after the transition are recorded, forming a set of state transition pairs containing timing information.

[0039] The set of state transition pairs is vectorized to construct a logical state vector representing the operational logic of the equipment. Using one-hot encoding or embedded vector techniques, each discrete state is mapped to a numerical vector of fixed dimension and concatenated in chronological order to generate a computable logical state vector that reflects the evolution trajectory of the equipment's behavioral logic within the process unit.

[0040] Through the above-mentioned protocol parsing, state mapping, transition identification and vectorized encoding processing methods, the semi-structured log data of the previous step is transformed into discrete state event sequences and computable logical state vectors that characterize the logical transitions of device operation. This achieves semantic unification and structured representation of heterogeneous device state data, providing a standardized logical input basis for subsequent cross-modal correlation analysis.

[0041] For example, in the PLC log of the condenser assembly station, Modbus TCP packets are parsed to extract data from holding registers 40001-40010. The original hexadecimal status code 0x00A1 is mapped to "clamping in place", and 0x00B2 is mapped to "welding start". When the status transitions from "clamping in place" to "welding start" at t=10.5s, the transition is recorded.<Clamp,Weld> The sequence is encoded into a 10-dimensional one-hot vector to generate a logical state vector [0,1,0...0], thereby realizing a digital representation of device behavior and significantly improving the accuracy of state recognition.

[0042] S1.4: Based on the temporal overlap relationship between the discrete state event sequence and the structured sensor data stream, perform cross-modal context association strength calculation to generate an abnormal event context association index set that describes the coupling degree between parameter fluctuations and state transitions within a specific time window before and after the abnormal event is triggered.

[0043] The discrete state event sequence generated by S1.3 and the structured sensor data stream slices obtained by S1.2 are acquired, and a synchronous analysis window is established with the trigger time of the abnormal event as the time reference. The discrete state event sequence is time-series scanned to locate all state transition points marked as warnings or errors, and their precise timestamps are extracted as anchor points for causal correlation analysis.

[0044] Based on the anchor timestamp, a local time window is constructed by retrospectively observing a preset pre-observation duration and extending the post-response duration, encompassing the steady-state characteristics before the anomaly and the transient response after it occurs. Within this local time window, sliding variance calculation is performed on the structured sensor data stream to quantify the severity of fluctuations in key physical quantities such as torque and current before and after the anomaly is triggered.

[0045] A mutual information algorithm is used to evaluate the nonlinear dependency between discrete state transitions and continuous parameter fluctuations. The calculated mutual information values ​​are normalized to generate a contextual correlation strength index characterizing the tightness of coupling between state transitions and parameter fluctuations. Morphological features are extracted from the parameter waveforms within a local time window to identify characteristic peaks and valleys that have a phase-locked relationship with the state transitions.

[0046] By calculating the cross-modal context association strength, the discrete events and continuous data from the previous step are transformed into a set of abnormal event context association indicators that describe the coupling degree of parameter fluctuations and state transitions before and after the triggering of abnormal events, thereby achieving deep fusion and feature alignment of multi-source heterogeneous data at the semantic level.

[0047] For example, at the condenser assembly station, the pre-observation time is set to 2 seconds, and the post-response time is set to 3 seconds. When the PLC reports an "insufficient fixture pressure" alarm ID, the system extracts pressure sensor data within 5 seconds before and after that moment. The standard deviation of the pressure value in the 2 seconds before the alarm is calculated to be 0.5 MPa, and it drops sharply to 0.1 MPa within 1 second after the alarm. The correlation between the alarm event and the pressure drop is calculated using the mutual information formula, yielding a mutual information value of 0.85. After normalization, the correlation strength index is 0.92, indicating a strong coupling between the alarm and the pressure anomaly. The system also records the falling edge slope of the pressure waveform at -2.0 MPa / s, storing it as a morphological feature in the index set. This process significantly improves the accuracy of anomaly attribution and avoids misjudgments caused by a single data source.

[0048] S1.5: The operation time sequence topology, discrete state event sequence and abnormal event context association index set of the independent process data slice set are fused and encapsulated to generate an original behavior pattern signature set containing operation sequence topology, state transition frequency distribution and abnormal event context association strength.

[0049] S2: Based on the original behavioral pattern signature set, generate a hash vector representation for each process unit, construct a process semantic fingerprint database, and transform the original behavioral pattern signature set into a process semantic fingerprint sequence. Specifically, this includes: S2.1: Based on the state transition frequency distribution data in the original behavior pattern signature set, the steady-state duration of key parameters of each process unit within the standard cycle window is statistically calculated to generate a steady-state duration proportion feature vector that characterizes the stability of process operation.

[0050] Receive the raw behavior pattern signature set output by S1.5, which contains the discrete state event sequence and structured sensor data stream of each process unit within the standard beat window.

[0051] The discrete state event sequence is timestamped to extract the entry and exit times of each logical state node, and a list of state dwell times within the process is constructed.

[0052] For structured sensor data streams, a dynamic steady-state judgment threshold is set. This threshold is determined based on three times the standard deviation of parameters from historical good product data and is used to identify the stable operating range of key process parameters.

[0053] The sensor sampling points are mapped to the corresponding time axis, and continuous time periods in which the parameter fluctuation amplitude is lower than the dynamic steady-state judgment threshold are selected and marked as the steady-state interval of the key parameters.

[0054] Calculate the total duration of the steady-state interval of the key parameters and perform a ratio calculation with the total duration of the standard beat window to generate the initial steady-state percentage value.

[0055] A weighted fusion strategy is adopted, which linearly combines the proportion of the main process state residence time in the state residence time list with the proportion of the initial steady state. The weight coefficient is determined based on the influence factor of each state on product quality.

[0056] The final steady-state duration proportion characteristic value is calculated using the following formula: ; Where R is the characteristic value of the proportion of steady-state duration, w1 and w2 are the state dwell weight and parameter steady-state weight, respectively, and T state T represents the total dwell time in the main process states. stable The key parameter is the total steady-state duration, T. cycle This is the total duration of the standard beat window.

[0057] The calculated steady-state duration proportion feature value is standardized by Z-Score to eliminate the dimensional differences between different processes and generate a dimensionless steady-state duration proportion feature vector.

[0058] Through the above statistical calculations and standardization processes, the original behavioral pattern signature from the previous step is transformed into a feature vector representing the steady-state duration ratio, which characterizes the stability of the process operation. This achieves a quantitative representation of the health status of the process and provides a highly discriminative stability indicator for the subsequent construction of a compact hash vector.

[0059] For example, at the condenser assembly station, the standard cycle time window T cycle Set to 120 seconds. Parse the PLC log to obtain the residence time T for the main process state "vacuum extraction". state The duration is 45 seconds. Monitoring current sensor data, a dynamic steady-state threshold is set at ±5% of the rated current, and the total length T of continuous time periods where current fluctuations are below this threshold is selected. stable The duration is 90 seconds. Weights w1 and w2 are set to 0.4 and 0.6 respectively. Substituting into the formula, R = 0.4 * (45 / 120) + 0.6 * (90 / 120) = 0.15 + 0.45 = 0.60. After standardizing this value, one dimension of the steady-state duration proportion feature vector is generated. If a batch experiences frequent interruptions in the "vacuum extraction" state due to a vacuum pump malfunction, T... state Reduced to 30 seconds, T stable If the time is reduced to 60 seconds, the R value drops to 0.45, which significantly reflects the decrease in process stability and thus produces obvious semantic deviations in subsequent fingerprint comparisons.

[0060] S2.2: Using the steady-state duration proportion feature vector and the operation sequence topology in the original behavior pattern signature set, the discrete behavior features are encoded and mapped to generate a composite feature descriptor containing temporal logic and state distribution information.

[0061] Receive the steady-state duration proportion feature vector generated by S2.1 and the operation sequence topology data in the original behavior pattern signature set output by S1.5.

[0062] The topology of the operation sequence is analyzed by directed graph traversal, and the frequency of transition paths and average dwell time between discrete state nodes in the process unit are extracted to construct the state transition adjacency matrix.

[0063] The state transition adjacency matrix is ​​flattened into a one-dimensional state transition distribution vector, which represents the temporal dependency of the internal logic execution of the process.

[0064] Z-Score normalization is applied to the steady-state duration proportion feature vector to eliminate numerical scale differences caused by different physical dimensions and generate normalized steady-state feature sub-vectors.

[0065] The normalized steady-state feature vector is concatenated with the one-dimensional state transition distribution vector to form a preliminary fused feature tensor containing static stability index and dynamic logic flow information.

[0066] By introducing process semantic weight coefficients, the key state node features in the preliminary fusion feature tensor are weighted and enhanced, highlighting the contribution of core process links that are sensitive to quality anomalies.

[0067] By using a weighted fusion processing method, the result of the previous step is transformed into a composite feature descriptor containing temporal logic and state distribution information, thereby realizing a unified semantic representation of multi-source heterogeneous behavioral features and providing a highly discriminative input basis for subsequent hash mapping.

[0068] S2.3: Based on the composite feature descriptor, perform dimensionality reduction and compression processing using the local sensitive hash algorithm to generate a compact hash vector representation with a fixed length and maintaining semantic similarity distance characteristics.

[0069] The algorithm receives the composite feature descriptor generated by S2.2, which contains temporal logic and state distribution information, and uses it as the input vector for the Locality Sensitive Hash (LSH) algorithm. The composite feature descriptor undergoes normalization preprocessing to eliminate interference from features of different dimensions on the hash mapping distance metric, ensuring a uniform weight distribution of features across all dimensions in the vector space. A random hyperplane projection matrix is ​​constructed, generated from random numbers following a standard normal distribution, with the number of rows equal to the fixed length of the target hash vector and the number of columns equal to the dimension of the composite feature descriptor. The normalized composite feature descriptor is multiplied by the random hyperplane projection matrix to obtain a projected low-dimensional real-valued vector. A sign function is applied to the projected real-valued vector, mapping elements greater than zero to 1 and elements less than or equal to zero to 0, thus generating a binary hash code sequence. A shift-XOR optimization mechanism is introduced, performing cyclic shifts and XOR operations on the initial binary hash code sequence to enhance the avalanche effect of the hash value, ensuring that small differences in input features lead to significant changes in the output hash vector. The Hamming distance between the generated hash vector and existing fingerprints in the historical database is calculated to verify its ability to maintain semantic similarity distance characteristics; that is, semantically similar process behavior patterns should produce hash vectors with small Hamming distances. Through locality-sensitive hashing dimensionality reduction and compression, the high-dimensional composite feature descriptor from the previous step is transformed into a compact hash vector representation with a fixed length that maintains semantic similarity distance characteristics, achieving the expected technical effect of millisecond-level similarity retrieval.

[0070] S2.4: Construct a key-value index structure based on the compact hash vector representation, and write the compact hash vector representation of each process unit into the storage unit to generate a process semantic fingerprint database that supports millisecond-level retrieval response.

[0071] It receives compact hash vector representations and corresponding unique identifiers of process units as input data sources for building the index structure.

[0072] Initialize a key-value-based in-memory database instance, configure the number of hash slots and conflict resolution strategy, and establish a cache space for storing semantic fingerprints of the process.

[0073] The unique identifier of the process unit is serialized into a standard string format and set as the primary key in the index structure to ensure the uniqueness and retrieval of the primary key within the global production batch.

[0074] The compact hash vector representation is converted into a binary byte stream or a fixed-length floating-point array format and set as a value in the index structure, preserving the vector's dimension information and numerical precision.

[0075] Perform atomic write operations to persist the mapping relationship between primary keys and values ​​to the hash table of the in-memory database, and use memory addressing mechanism to eliminate disk I / O latency.

[0076] Attach timestamp metadata and workstation source tags to each written fingerprint data, and build a composite index field to support multi-dimensional fast filtering based on time window and physical location.

[0077] An inverted index auxiliary structure is established to map the key feature bits of the hash vector to the corresponding process ID list, thereby optimizing the candidate set screening efficiency during large-scale similarity comparison.

[0078] Through the key-value index construction and metadata enhancement processing methods described above, the discrete and compact hash vector generated in the previous step is transformed into a process semantic fingerprint database that supports millisecond-level random reading and range query, thereby achieving the expected technical effect of efficient organization and real-time retrieval of massive process fingerprint data.

[0079] S2.5: Call the compact hash vector representation in the process semantic fingerprint library, and serialize and assemble the compact hash vector representation of each process unit according to the production cycle order to generate a standardized process semantic fingerprint sequence as the final output.

[0080] The compact hash vector representations of each key process unit in the current production batch are read from the process semantic fingerprint database and used as the original data input source for serialization assembly.

[0081] Based on the standard process routing table issued by the production line main control system, extract the process flow topology sequence corresponding to the current air conditioner top-mounted integrated unit model, and determine the logical order of each process node on the physical assembly line.

[0082] The compact hash vector representation is key-value matched with the process flow topology sequence, and the discrete hash vectors are linearly sorted according to the increasing timestamp and process dependency.

[0083] Perform integrity verification on the sorted hash vector sequence to detect whether there are missing nodes in the process fingerprint due to sensor failure or communication packet loss.

[0084] If a missing node is detected, a null placeholder is inserted or a local re-acquisition command is triggered based on the time interval threshold of the fingerprints of adjacent processes to ensure the continuity of the sequence in the time dimension.

[0085] The sorted and complete set of compact hash vectors is encapsulated into a fixed-length one-dimensional array structure, with each array element corresponding to the semantic fingerprint of a specific process unit.

[0086] Add metadata header information such as batch identifier, production time window start stamp and end stamp to the one-dimensional array to form a standard data packet with a unique index identifier.

[0087] Through the above-mentioned serialization assembly process, the compact hash vectors of each process stored in the previous step are transformed into a standardized process semantic fingerprint sequence arranged according to the process time sequence. This realizes the structured mapping of multi-source heterogeneous data to a unified time-series semantic index, providing an ordered and complete data foundation for the subsequent topological traversal of the causal skeleton graph.

[0088] S3: Utilize the process semantic fingerprint sequence to statistically verify the cross-process influence relationship, mapping the discrete process semantic fingerprint sequence into a basic causal skeleton graph structure with nodes representing process semantic fingerprints and edges representing lag correlations. Specifically, this includes: S3.1: Obtain the standardized process semantic fingerprint sequence corresponding to the historical good product production batch, and perform multi-station time sequence reorganization processing on the standardized process semantic fingerprint sequence based on the timestamp alignment mechanism to generate a good product process semantic fingerprint time sequence matrix arranged in the production cycle order.

[0089] Standardized process semantic fingerprint sequences corresponding to historical good product production batches are obtained as initial data input for constructing the basic causal skeleton graph. Based on the precise timestamps recorded by the data acquisition terminals at each workstation, global temporal alignment processing is performed on the multi-source heterogeneous fingerprint sequences to eliminate time deviations caused by network transmission jitter or edge computing latency. According to the standard process cycle parameters of the air conditioner rooftop integrated production line, a fixed-width sliding time window is set to divide the continuous production process into several discrete process unit time slices. Within each time slice, the process semantic fingerprint of the corresponding workstation is extracted and logically recombined according to the sequence of the physical process flow. Through matrix encapsulation technology, the recombined fingerprint data is mapped into a two-dimensional good product process semantic fingerprint temporal matrix, where the row vectors represent the complete process link of a specific batch, and the column vectors represent the fingerprint evolution trajectory of a specific process under different batches. Through the above processing method, the discrete fingerprint sequences generated in the previous step are transformed into structured temporal matrix data with strict temporal constraints and batch associations, realizing the standardized representation of the historical good product production status and providing a unified and unbiased data benchmark for subsequent statistical verification of cross-process nonlinear dependency strength.

[0090] S3.2: Based on the temporal matrix of the semantic fingerprint of the good product process, the nonlinear dependency strength between the process semantic fingerprints of any two different process units is calculated using the mutual information estimation algorithm to generate a candidate causal association set containing all potential association pairs and their dependency strength values.

[0091] Receive the good product process semantic fingerprint time series matrix generated by S3.1. The row vectors of this matrix correspond to the compact hash fingerprints of specific process units, and the column vectors correspond to the time series indices sorted by production cycle.

[0092] Traverse the row vector sequences of any two different process units in the matrix to construct the paired fingerprint time series combination to be analyzed, ensuring coverage of all potential upstream and downstream and cross-workstation related pairs.

[0093] For each fingerprint sequence, the kernel density estimation method is used to calculate the marginal probability distribution function of the fingerprint feature values ​​of each process, and the discrete hash vector is mapped to a continuous probability density space to eliminate the influence of quantization noise.

[0094] Based on the edge probability distribution, the joint probability density function of the fingerprint sequences of the two processes is jointly calculated. By statistically analyzing the co-occurrence frequency and time lag offset, a two-dimensional joint distribution histogram is constructed and smoothed.

[0095] The nonlinear dependency strength between two process fingerprint sequences is calculated using the mutual information definition formula, quantifying the amount of information contained in the behavioral pattern of one process by the behavioral pattern of another process.

[0096] The mutual information values ​​of all process pairs are normalized to eliminate dimensional bias caused by differences in fingerprint dimensions and generate standardized nonlinear dependency strength coefficients.

[0097] Select process pairs with a dependency strength coefficient greater than a preset threshold, record their corresponding process identifier, lag step size and strength value, and form a candidate causal association set.

[0098] By using kernel density estimation and mutual information calculation, the time series matrix from the previous step is transformed into a set of candidate causal relationships that characterize the nonlinear coupling strength between processes, thus achieving preliminary quantitative screening of potential causal relationships.

[0099] S3.3: Based on the candidate causal association set, the Granger causality test method is applied to verify the time-delay significance of the nonlinear dependency strength, so as to screen out statistically significant lagged causal relationships and generate a statistically verified list of cross-process influence relationships.

[0100] Receive the candidate causal association set generated by S3.2. This set contains the nonlinear dependency strength value and the corresponding time lag interval between the semantic fingerprints of any two different process units.

[0101] For each pair of process fingerprint sequences in the candidate causal association set, a vector autoregressive (VAR) model is constructed to quantify the dynamic linear dependencies between time series. A maximum lag order is set, and the optimal lag order is determined based on the Akaike Information Content Criterion (AIC) to ensure that the model can capture short-term dynamics while avoiding overfitting.

[0102] Based on the constructed VAR model, Granger causality tests were performed to verify whether changes in upstream process fingerprints significantly predict changes in downstream process fingerprints. The sum of squared residuals for the constrained and unconstrained models was calculated, and statistical measures were used to evaluate the significance of the improvement in the prediction accuracy of downstream variables after introducing upstream variables.

[0103] The calculated statistic is compared with the critical value at the preset significance level. If the statistic is greater than the critical value and the lag order is less than 0.05, the null hypothesis is rejected, and it is determined that there is a statistically significant Granger causal relationship between the upstream process fingerprint and the downstream process fingerprint.

[0104] Record the process pairs that pass the significance test and their corresponding optimal lag time steps, and remove spurious correlations that fail the test. Integrate all validated process pairs to form a statistically validated list of cross-process influence relationships with directionality and time lag properties.

[0105] By using the Granger causality test, the nonlinear dependency candidate set from the previous step is transformed into a list of directional causal connections with statistical significance, achieving a logical leap from correlation to causality and providing rigorous data support for constructing a high-confidence basic causal skeleton graph.

[0106] S3.4: Based on the statistically verified list of cross-process influence relationships, construct a basic causal skeleton graph with process semantic fingerprints as nodes and lag correlations as directed edges, and perform a weight stripping operation on the basic causal skeleton graph to generate a basic causal skeleton graph structure that retains only the topological connection structure without fixing the specific weight values.

[0107] Receive a statistically validated list of cross-process influence relationships, which includes the source process fingerprint ID, the target process fingerprint ID, and the corresponding lag correlation confidence scores.

[0108] Traverse each record in the influence relationship list and extract the source process fingerprint ID and the target process fingerprint ID as unique identifiers for the graph nodes.

[0109] Initialize an empty directed graph data structure in memory, instantiate all extracted unique process fingerprint IDs as node objects in the graph, and store the corresponding compact hash vector representation of each node object.

[0110] For each source-target process fingerprint in the list, create a directed edge in the directed graph from the source node to the target node to represent the potential causal flow.

[0111] Perform a weight stripping operation to remove the hysteresis correlation confidence values ​​and other quantized weight parameters attached to the directed edges, retaining only the existence state and direction attributes of the edges.

[0112] Perform a topological connectivity check on the generated directed graph, remove all isolated nodes with zero in-degree and zero out-degree, and ensure that the basic causal skeleton graph contains only process units with potential relationships.

[0113] The processed directed graph structure is serialized and stored to form a basic causal skeleton graph structure that retains only the topological connections without fixing the specific weight values.

[0114] By using the above processing method, the statistical verification results of the previous step are transformed into lightweight topological structure data, achieving the expected technical effect of fast traversal without loading a large weight matrix in the subsequent real-time attribution process.

[0115] like Figure 2As shown, S4: In response to the quality anomaly signal detected in real time, the current abnormal process fingerprint is used as the query starting point. A restricted depth-first traversal is performed on the basic causal skeleton graph, and adjacent candidate process semantic fingerprints are dynamically loaded. The semantic deviation between the current abnormal process fingerprint and the candidate process semantic fingerprints is calculated, and a dynamically activated edge set is generated. Specifically, this includes: S4.1: Obtain the quality anomaly signal detected in real time and the corresponding abnormal batch identifier, extract the end process fingerprint of the current abnormal batch from the process semantic fingerprint database based on the abnormal batch identifier, and set the end process fingerprint as the initial query starting node of the restricted depth-first traversal.

[0116] The system receives real-time quality anomaly signals triggered by online quality inspection terminals and parses the unique batch tracking identifier and anomaly occurrence timestamp carried in the signal. Based on the unique batch tracking identifier, it performs an exact match retrieval in the key-value index structure of the process semantic fingerprint database to locate the complete process semantic fingerprint data block corresponding to the current anomaly batch in the final process of the production process. It extracts a compact hash vector representation from the data block, which includes the operation sequence topology, state transition frequency, and anomaly context association strength information of the final process within the standard beat window. The extracted final process compact hash vector is encapsulated into a graph traversal query object, assigning it attribute labels with an initial traversal depth of zero and an initial confidence level of full. This graph traversal query object is set as the initial query starting node of the restricted depth-first traversal algorithm on the basic causal skeleton graph, establishing the logical starting position for reverse tracing. Through the above retrieval and encapsulation processing, discrete anomaly signals are transformed into initial query starting nodes with graph theory topological meaning, achieving a precise mapping from physical anomaly events to the semantic graph search space, providing a definite data anchor point for rapid location of subsequent upstream processes.

[0117] For example, when the airtightness testing station detects a leakage rate exceeding the standard, the system parses the batch ID as BN20231025-089. The system retrieves the fingerprint of the last process corresponding to this ID from the fingerprint database, obtaining a compact hash vector H_T99=[0x1A, 0xFF, ..., 0x3B] with a length of 128 bits. The system encapsulates H_T99 into a query object Node_Start, setting depth=0. This object is input as the starting point to the causal skeleton graph traversal engine, ensuring that subsequent backtracking only expands on upstream nodes directly related to the end state of this specific batch, avoiding the computational delay caused by full graph scanning, and significantly improving the real-time performance of the attribution response.

[0118] S4.2: Based on the topological connection relationship of the initial query starting node in the basic causal skeleton graph, locate the upstream adjacent process node directly connected to the initial query starting node, and dynamically load the latest batch process semantic fingerprint sequence corresponding to the upstream adjacent process node from the process semantic fingerprint database according to the preset recent batch quantity threshold.

[0119] Receive the initial query starting node identifier and basic causal skeleton graph topology data output by S4.1 as the input conditions for this step.

[0120] Analyze the set of upstream adjacent process nodes in the basic causal skeleton graph that are connected to the initial query starting node by directed edges, and extract the process unique identifier of each upstream node.

[0121] Based on the preset threshold N for the number of most recent batches, a fingerprint retrieval request instruction is constructed for each upstream adjacent process node, where the value of N is dynamically adjusted according to the stability of the production line cycle time.

[0122] Send a batch retrieval command to the process semantic fingerprint database, specifying the retrieval key as the identifier of the upstream adjacent process node, and the retrieval time range as the time window of N production cycles backward from the current abnormal batch time.

[0123] The process semantic fingerprint database performs an index lookup operation to locate the compact hash vector representation sequence generated by the corresponding process node within a specified time window.

[0124] Extract the retrieved compact hash vector representation, arrange it in reverse order of timestamp, and generate the semantic fingerprint sequence of the most recent batch of processes containing N consecutive batches of data.

[0125] The integrity of the generated latest batch of process semantic fingerprint sequences is checked to ensure that the sequence length is equal to the preset threshold N and that the dimension of each fingerprint vector is consistent.

[0126] By using topology-based targeted retrieval and time window slicing, the query starting point of the previous step is transformed into a set of recent behavioral pattern data of the upstream process, achieving low-latency local context data loading and providing an accurate data benchmark for subsequent semantic deviation calculation.

[0127] S4.3: Perform vector space mapping on the final process fingerprint contained in the initial query starting node and each candidate process semantic fingerprint in the most recent batch process semantic fingerprint sequence, and use the Hamming distance algorithm to calculate the discrete feature difference value between the final process fingerprint and each candidate process semantic fingerprint to generate the original semantic deviation value set.

[0128] The fingerprint vector of the last process contained in the initial query starting node and the latest batch of process semantic fingerprint sequence of the upstream adjacent process nodes dynamically loaded from the process semantic fingerprint database are obtained as input data objects for deviation calculation.

[0129] A bit alignment operation is performed on the final process fingerprint and the semantic fingerprint of each candidate process in the sequence to ensure that the two fixed-length compact hash vectors correspond strictly in the bit index, thus eliminating the calculation error caused by storage offset.

[0130] The two aligned binary vectors are processed by bitwise XOR logic operation to generate an intermediate difference vector that reflects the difference in each bit. Bits with a value of 1 indicate that the characteristics are inconsistent, and bits with a value of 0 indicate that the characteristics are consistent.

[0131] The total number of bits with a value of 1 in the intermediate difference vector is counted. This count directly represents the Hamming distance between the semantic fingerprints of the two processes in the discrete behavioral feature space, serving as the quantification basis for the original semantic bias.

[0132] The calculated Hamming distance value is stored in a temporary cache and mapped to the unique identifier of the semantic fingerprint of the current candidate process to form the original semantic deviation data item in the form of key-value pairs.

[0133] Iterate through all the semantic fingerprints of the most recent batch of processes corresponding to the upstream adjacent process nodes, and repeat the bit alignment, XOR operation and bit counting process described above until the deviation calculation of all candidate fingerprints in the sequence is completed.

[0134] Aggregate all generated raw semantic deviation data items to construct a set of raw semantic deviation values ​​containing multiple deviation values ​​and their corresponding fingerprint identifiers, which will be used for subsequent threshold filtering.

[0135] By using the Hamming distance algorithm to perform discrete feature difference quantization on the semantic fingerprint of the process, the multidimensional fingerprint sequence loaded in the previous step is transformed into a scalarized set of original semantic deviation values. This achieves efficient dimensionality reduction mapping from high-dimensional data to low-dimensional similarity indicators in the anomaly attribution process, significantly reducing real-time computing load and improving response speed.

[0136] For example, the semantic fingerprint length of a process is set to 128 bits. The fingerprint of the final process is a binary string A, and the fingerprint of a candidate fingerprint in a certain batch upstream is a binary string B. An XOR operation is performed on A to obtain the difference string C. The number of 1s in C is counted to be 15, i.e., the Hamming distance is 15. If the preset deviation threshold is 10, then 15 is greater than 10, and the candidate fingerprint is marked as an abnormal associated node. The deviation is calculated one by one for the 10 most recent batches of fingerprints upstream, resulting in the deviation set {15, 8, 12, 9, 20, ...}. This process takes less than 1 millisecond, significantly improving real-time attribution efficiency.

[0137] S4.4: Based on the comparison and judgment between the original set of semantic deviation values ​​and the preset semantic deviation threshold, abnormal associated nodes with a value greater than the semantic deviation threshold are selected, and the topological connection between the abnormal associated nodes and the initial query starting node is marked as a high-confidence potential causal edge to generate a set of edges to be verified.

[0138] Receive the raw semantic deviation value set generated by step S4.3, which contains the Hamming distance values ​​between the fingerprint of the end process of the current abnormal batch and the candidate fingerprints of the nearest N batches of the upstream adjacent process.

[0139] Each deviation value in the original set of semantic deviation values ​​is traversed and compared with the semantic deviation threshold preset by the system. The threshold is set based on the principle of three times the standard deviation of the historical good product fingerprint distribution and is used to define the boundary between normal process fluctuations and abnormal behavior patterns.

[0140] Identify and filter all deviation values ​​greater than the semantic deviation threshold, and mark the corresponding candidate process semantic fingerprints as abnormal association nodes, indicating that the upstream process has shown significant feature variations that deviate from the standard behavior pattern in the corresponding batch.

[0141] Obtain the topological edge information of the connection between the initial query starting node and the above-mentioned abnormal associated nodes in the basic causal skeleton graph, and mark the topological edge as a high-confidence potential causal edge, indicating that the connection has a very high probability of causal transmission under the current abnormal scenario.

[0142] All marked high-confidence potential causal edges and their associated abnormal nodes are encapsulated into a set of edges to be verified, forming the next level of search space for restricted depth-first traversal.

[0143] By using threshold comparison screening and topological labeling, the discrete feature difference values ​​calculated in the previous step are transformed into a set of edge data to be verified with causal orientation. This achieves the expected technical effect of quickly locking down high-probability anomaly sources from a massive batch of candidates, and significantly reduces the computational load of subsequent causal strength reassessment.

[0144] For example, let's assume a preset semantic deviation threshold of 12. The calculated Hamming distances for the five most recent batches of upstream welding stations are 8, 15, 9, 18, and 7, respectively. A comparison reveals that 15 and 18 are greater than the threshold of 12. The system marks the fingerprints of welding stations with batch indices 2 and 4 as anomalous association nodes. In the basic causal skeleton graph, the edges connecting the airtightness testing station (the current starting point) to these two welding station nodes are marked as high-confidence potential causal edges. The final generated set of edges to be verified contains two edges, pointing to the fingerprint nodes of welding station batches 2 and 4 respectively, excluding the other three normally fluctuating batches. This allows subsequent recursive traversal to focus only on these two nodes, significantly reducing the search scope.

[0145] S4.5: Based on the set of edges to be verified, perform a recursive restricted depth-first traversal operation, update the latest abnormal associated node in the traversal process to the new current query starting node, and repeat the process of loading adjacent nodes and calculating deviation until the traversal depth reaches the maximum level limit or no new high-confidence potential causal edges are generated. Finally, output a set of dynamically activated edges consisting of all marked high-confidence potential causal edges.

[0146] Receive the set of edges to be verified generated by S4.4, and extract the downstream node of each edge in the set as the new current query starting node.

[0147] Read the adjacency list structure of the basic causal skeleton graph and retrieve the list of upstream process nodes that have a direct topological connection with the current query starting node.

[0148] Based on the preset threshold N for the number of most recent batches, the process semantic fingerprint sequences of the most recent N batches corresponding to each upstream process node are concurrently loaded from the process semantic fingerprint database.

[0149] Align the semantic fingerprint of the current query starting node with the loaded upstream candidate fingerprints through vector space mapping.

[0150] The Hamming distance algorithm is used to calculate the discrete feature difference values ​​between the current fingerprint and each upstream candidate fingerprint, and to generate the original set of semantic deviation values.

[0151] A recursive traversal mechanism is adopted to mark upstream nodes with deviation exceeding a preset threshold as new high-confidence potential causal edges and add them to the set of dynamically activated edges.

[0152] Update the traversal depth counter to determine if the current depth has reached the maximum level limit L. max .

[0153] If the limit is not reached and a new high-confidence potential causal edge exists, the latest abnormal associated node is updated as the query starting point for the next round of traversal, and the process of loading adjacent nodes and calculating deviation is repeated.

[0154] If the limit is reached or no new high-confidence edges are generated, the recursive traversal operation is terminated.

[0155] Summarize all high-confidence potential causal edges marked in the traversal path and construct a dynamic set of activated edges containing temporal dependencies.

[0156] By employing a restricted depth-first traversal and dynamic deviation filtering method, discrete edges to be verified are transformed into a set of dynamically activated edges with a hierarchical structure, enabling rapid backtracking and location of the anomaly source.

[0157] For example, set the maximum traversal depth L. max The threshold for the most recent batch number is 5. The current query starting point is fingerprint T99 at the airtightness testing station, with a Hamming distance threshold of 0.15. The system loads the 5 most recent batches of fingerprints from the upstream welding station W07. The Hamming distance between the 3rd batch of fingerprints and T99 is calculated to be 0.12, which is less than the threshold, so it is determined to be a high-confidence edge and added to the set. Continuing upstream to the screw fastening station F12, the system loads its 5 most recent batches of fingerprints. The Hamming distance between the 1st batch of fingerprints and the 3rd batch of fingerprints from W07 is calculated to be 0.08, which is less than the threshold, so it is marked as a high-confidence edge. The search stops when the traversal depth reaches 3 layers. The final output contains a dynamically activated edge set [T99-W07, W07-F12], significantly shortening the attribution path search time.

[0158] S5: Based on the shared discrete behavioral features in the dynamically activated edge set, perform local causal strength re-evaluation and generate a dynamic causal graph with short-term memory factors, transforming it into a dynamic causal graph link. Specifically, this includes: S5.1: Perform discrete behavioral feature intersection operation on the semantic fingerprint nodes of upstream and downstream processes in the dynamically activated edge set, and extract shared discrete behavioral feature vectors including the probability of synchronous occurrence and the propagation attenuation coefficient to construct a feature input benchmark set for local causal strength re-evaluation.

[0159] Receive semantic fingerprint nodes of upstream and downstream processes in the dynamically activated edge set, and parse the discrete behavioral feature vectors contained in each node. The feature vectors cover the PLC status code sequence, alarm ID frequency distribution and operation timing topology identifier.

[0160] Perform set intersection operation on the discrete behavioral feature vectors of upstream and downstream processes to identify common behavioral patterns that occur synchronously within the time window and generate a shared discrete behavioral feature candidate set.

[0161] Based on a shared set of discrete behavioral features, the frequency of simultaneous triggering of specific discrete events in upstream and downstream processes is statistically analyzed, and a synchronous occurrence probability index is calculated to quantify the tightness of cross-process behavioral coupling.

[0162] For shared discrete events with temporal sequence relationships, the propagation time interval data between adjacent processes is extracted, and a time decay model is constructed to calculate the propagation attenuation coefficient, characterizing the transmission loss characteristics of abnormal signals in the process chain.

[0163] The probability index of synchronous occurrence and the propagation attenuation coefficient are concatenated and normalized to construct a shared discrete behavioral feature vector containing coupling strength and temporal dependency information, which serves as the feature input benchmark set for local causal strength re-evaluation.

[0164] Through the above processing method, the dynamic activation edges of the previous step are transformed into shared discrete behavioral feature vectors with quantized coupling characteristics, realizing an accurate mapping from topological connection to semantic association strength, and providing high-dimensional feature support for subsequent Bayesian confidence updates.

[0165] For example, in an integrated air conditioner production line, the fingerprint at the upstream screw fastening station F12 contains a torque fluctuation pattern, while the fingerprint at the downstream pipe welding station W07 contains a current fluctuation pattern. The system extracts the shared discrete features of both: 'PLC alarm code E03' and 'scanning interval abrupt change S2'. Statistics from nearly 100 batches show that E03 occurs synchronously 85 times in both upstream and downstream processes, with a calculated synchronous occurrence probability of 0.85. Further analysis of the E03 trigger time difference reveals an average lag time of 2.5 seconds and a standard deviation of 0.3 seconds, based on an exponential decay model. The propagation attenuation coefficient was calculated, where λ is the attenuation constant of 0.4 and Δt is the lag time of 2.5, yielding an attenuation coefficient α of approximately 0.37. Finally, a shared feature vector [0.85, 0.37] was generated, significantly improving the discriminative power of causal association discrimination and effectively filtering out sporadic noise without temporal correlation.

[0166] S5.2: Based on the shared discrete behavior feature vector, the initial topological connection of each edge in the dynamically activated edge set is conditionally probabilistically calculated using the Bayesian network inference algorithm to generate a real-time posterior confidence value that characterizes the credibility of causal association under the current abnormal batch.

[0167] The system receives a shared discrete behavior feature vector output from S5.1. This vector contains the PLC alarm code sequence that appears synchronously in the upstream and downstream process fingerprints, the barcode interval mutation marker, and the parameter steady-state deviation marker. A local Bayesian network structure is constructed with the currently dynamically activated edge as the inference unit. The abnormal behavior features in the upstream process fingerprint are set as parent node variables, and the response features in the downstream process fingerprint are set as child node variables, thus establishing the topological dependency relationship for causal inference.

[0168] Extract the joint distribution statistics of discrete behavioral features from historical good batches and known abnormal batches, and calculate the conditional prior probability of the child node state occurring given the parent node state. For the current real-time abnormal batch, obtain the specific observed states of the parent and child nodes, and substitute them into the Bayesian inference formula for posterior probability updates.

[0169] The real-time posterior confidence level is calculated using the following formula: ; Wherein, P(C|E) is the posterior confidence of the existence of upstream causal feature C under the condition that downstream abnormal feature E is observed; P(E|C) is the likelihood function, which represents the probability that downstream feature E occurs when upstream feature C exists, and is obtained by normalizing historical statistical frequency; P(C) is the prior probability, which reflects the basic occurrence frequency of feature C in the overall production process; and P(E) is the evidence factor, which is the marginal probability of feature E in all historical data.

[0170] The calculated posterior confidence scores are normalized and mapped to a confidence interval of 0 to 1, generating a real-time posterior confidence score that characterizes the causal relationship strength of the dynamically activated edge in the current abnormal batch. This score quantifies the degree of certainty that upstream process behavior leads to downstream quality anomalies under the current specific combination of operating conditions.

[0171] By using a Bayesian network inference algorithm, the shared discrete behavioral features extracted in the previous step are transformed into real-time posterior confidence values, enabling probabilistic quantitative evaluation of the causal strength of dynamically activated edges. This significantly improves the robustness and interpretability of attribution results in complex noisy environments.

[0172] For example, in the causal chain analysis between the air conditioner condenser assembly station and the airtightness testing station, shared discrete features include "torque wrench alarm code A03" and "airtightness test pressure drop indicator B12". Historical data shows that P(B12|A03) is 0.85, P(A03) is 0.02, and P(B12) is 0.05. When A03 and B12 are detected simultaneously in real time, the posterior confidence level P(A03|B12) = (0.85 × 0.02) / 0.05 = 0.34 is calculated using the formula. If the shared features of the other path are "scanning delay" and "missing installation detection", the calculated posterior confidence level is 0.12. Based on this, the system determines that the causal confidence level of the torque anomaly path is higher, and prioritizes it for investigation, effectively narrowing down the fault location range.

[0173] S5.3: Based on the real-time posterior confidence value, call the historical abnormal event recurrence frequency statistics module to obtain the recurrence frequency data of the corresponding edge in the last M abnormal events, and perform weighted fusion processing on the recurrence frequency data and the real-time posterior confidence value to generate a comprehensive causal strength index including timeliness weight.

[0174] The system receives the real-time posterior confidence score from step S5.2 and uses it as the basic metric for the credibility of causal association in the current abnormal batch. It then calls the historical abnormal event recurrence frequency statistics module, using each edge in the dynamically activated edge set as the index key to retrieve the recurrence frequency data of that edge within the most recent M historical abnormal event windows. The retrieved recurrence frequency data is normalized to eliminate counting bias caused by differences in the cycle times of different processes, generating a standardized historical recurrence frequency factor. A comprehensive causal strength calculation model incorporating timeliness weights is constructed, and the real-time posterior confidence score is weighted and fused with the standardized historical recurrence frequency factor. The comprehensive causal strength index is calculated using the following formula: ; Wherein, I is the comprehensive causal strength index, P is the real-time posterior confidence value, F is the standardized historical recurrence frequency factor, and α and β are the real-time confidence weight coefficient and the historical recurrence weight coefficient, respectively, satisfying α+β=1. Based on the stability requirements of the production environment, the ratio of the weight coefficients α and β is dynamically adjusted. When the production line conditions fluctuate drastically, the value of α is increased to enhance real-time response sensitivity, and the value of β is increased during steady-state production to utilize historical statistical patterns to suppress occasional noise. The calculated comprehensive causal strength index is mapped to the [0,1] interval to form a physically interpretable quantitative score. Through the above weighted fusion processing method, the single probability confidence value from the previous step is transformed into a comprehensive causal strength index that incorporates historical statistical patterns, achieving effective filtering of occasional interference and strengthening of stable causal links, significantly improving the robustness of the attribution results.

[0175] For example, the historical window size M is set to 50 batches of abnormal events. For a specific edge in the dynamically activated edge set connecting the "condenser assembly station" and the "air tightness test station", the real-time posterior confidence level P output by S5.2 is 0.75. The statistics module found that this edge reproduced 40 times in the last 50 anomalies, and the historical recurrence frequency factor F after normalization is 0.8. Based on the current production line being in a steady-state production mode, the real-time confidence weight coefficient α is set to 0.4, and the historical recurrence weight coefficient β is set to 0.6. Substituting into the formula, we calculate: I = 0.4 × 0.75 + 0.6 × 0.8 = 0.3 + 0.48 = 0.78. This comprehensive causal strength index of 0.78 is higher than the confidence level of 0.75 relying solely on real-time data, indicating that the causal relationship has high historical stability and is marked by the system as a high-priority attribution path, effectively avoiding misjudgments caused by single sensor fluctuations.

[0176] S5.4: Based on the comprehensive causal strength index, perform a short-term memory factor mapping function operation to convert the comprehensive causal strength index into a short-term memory factor value that characterizes the stability of the causal chain, so as to quantify the degree of continuous effectiveness of the causal connection in the continuous production process.

[0177] Receive the comprehensive causal strength index generated by S5.3, which integrates real-time posterior confidence and historical recurrence frequency, as the input benchmark for calculating the short-term memory factor.

[0178] A nonlinear mapping function is constructed to transform the comprehensive causal strength index into a short-term memory factor value that characterizes the stability of the causal chain. This function needs to have saturation characteristics to suppress extreme value fluctuations.

[0179] A short-term memory factor calculation model is defined, and the Sigmoid activation function is used to normalize the comprehensive causal strength index to ensure that the output value range is strictly limited to 0 to 1.

[0180] The short-term memory factor is calculated using the following formula: ; Where M is the short-term memory factor value, I is the comprehensive causal strength index, k is the gain coefficient used to control the steepness of the mapping curve, θ is the activation threshold, and e is the natural constant.

[0181] The initial value of the gain coefficient k is set to 5.0 to avoid gradient vanishing while ensuring discriminative power. The activation threshold θ is set to 0.6, which corresponds to a baseline of moderate causal correlation.

[0182] Substitute the comprehensive causal strength index I into the formula, perform exponential and division operations, and obtain the corresponding short-term memory factor M.

[0183] The calculated short-term memory factor M is quantized and encoded, retaining four decimal places of precision to adapt to the weight storage format of subsequent graph links.

[0184] By using the Sigmoid mapping method, the comprehensive causal strength index of the previous step is transformed into a standardized short-term memory factor value, which realizes the quantitative characterization of the continuous effectiveness of causal connections in the continuous production process and enhances the robustness of the attribution model to incidental noise.

[0185] For example, in the attribution scenario of an airtightness test anomaly on an air conditioning unit production line, the calculated comprehensive causal strength index I of the dynamically activated edge between a connecting screw tightening station and a pipe welding station is 0.75. The gain coefficient k is set to 5.0, and the activation threshold θ is set to 0.6. Substituting into the formula to calculate the exponent: -5.0 * (0.75 - 0.6) = -0.75. e raised to the power of -0.75 is approximately 0.4724. The denominator is 1 + 0.4724 = 1.4724. The short-term memory factor M = 1 / 1.4724 ≈ 0.6792. This value indicates that the causal connection has high stability and is not considered an occasional interference. If the composite causality strength index I of the other edge is 0.55, then the exponential part is -5.0 * (0.55 - 0.6) = 0.25, e raised to the power of 0.25 is approximately 1.2840, the denominator is 2.2840, and M ≈ 0.4378. This lower memory factor value will be identified as a low-confidence edge and weakened in step S6, thereby significantly improving the accuracy of quality tracing.

[0186] S5.5: The edge attributes in the dynamically activated edge set are enhanced and labeled using the short-term memory factor values. The edges with short-term memory factor values ​​are recombined into dynamic causal graph links with recurrence frequency weighting attributes to output the final attribution path structure that supports interpretability tracing.

[0187] The system receives the short-term memory factor (STM) value set and corresponding dynamically activated edge set generated by S5.4, and injects the STM value as the core enhancement parameter for edge weights into the graph data structure. For each directed edge in the dynamically activated edge set, an attribute expansion operation is performed, adding a recurrence frequency weighted field to the original topology connection information. This field directly maps to the quantized value of the STM factor. A linear interpolation algorithm is used to normalize the STM factor, ensuring that the weight values ​​of all edges are distributed within a uniform confidence interval, eliminating evaluation bias caused by differences in dimensions between different processes. A dynamic causal graph link with time decay characteristics is constructed, connecting the weighted edges according to the reverse or sequential logic of the process flow to form a complete causal path chain with intensity indicators. Semantic labels are bound to the nodes in the link, associating the unique identifier of the process semantic fingerprint with the corresponding physical workstation name and anomaly feature description for storage, ensuring the readability of the path. Through the above processing method, the discrete causal intensity index calculated in the previous step is transformed into structured dynamic causal graph link data with recurrence frequency weighting attributes, realizing the transformation from abstract statistical values ​​to traceable and interpretable quality attribution path structures, and providing a data foundation with clear weighting basis for subsequent removal of occasional interference.

[0188] For example, in the anomaly attribution scenario of air tightness testing on an air conditioning unit production line, the dynamically activated edge set contains three key edges: [screw tightening → pipe welding], [pipe welding → refrigerant charging], and [refrigerant charging → air tightness testing]. The original short-term memory factors calculated by S5.4 are 0.85, 0.62, and 0.91, respectively. The system first normalizes these three values, mapping them to the [0,1] interval while maintaining their relative size. Subsequently, the system writes 0.85, 0.62, and 0.91 into the weight attribute field of the corresponding edge and marks the edge status as active. At the same time, the semantic fingerprint ID (such as F12, W07, C03, T99) and physical meaning (torque attenuation, current fluctuation, flow abnormality, pressure leakage) of each node are extracted and encapsulated into the node metadata attribute. The final output dynamic causal graph link is a JSON object or a graph database subgraph, where each edge not only contains the source node and target node IDs but also carries a specific recurrence frequency weighted value. This structure allows downstream steps to directly identify [pipe welding → refrigerant charging] as a possible incidental interference based on a low weight of 0.62, while a high weight of 0.91 confirms [refrigerant charging → airtightness test] as a strong causal relationship, significantly improving the structuring of the attribution path and the accuracy of subsequent processing.

[0189] like Figure 3As shown, S6: Based on the short-term memory factor threshold judgment result of the dynamic causal graph link, weaken the occasional interference edge connections below the set threshold and output the process semantic chain composed of process semantic fingerprint nodes and dynamic confidence scores, thus parsing the dynamic causal graph link into the final quality problem attribution path. Specifically, this includes: S6.1: Obtain the recurrence frequency weighted attribute and historical confidence data of each edge in the dynamic causal graph link, and normalize the recurrence frequency weighted attribute based on the sliding time window statistical mechanism to generate a set of short-term memory factor values ​​representing causal stability.

[0190] The recurrence frequency weighted attributes and historical confidence data of each edge are extracted from the dynamic causal graph link to construct a raw statistical data set containing time series dimensions.

[0191] A sliding time window mechanism is applied to the original statistical data set, with the window length set to the time span of nearly N production batches, in order to extract a local data subset that is representative of the timeliness.

[0192] Calculate the maximum and minimum recurrence frequencies of each edge within the sliding window to determine the range of numerical distribution within the current statistical period, providing a baseline boundary for normalization processing.

[0193] The range standardization method is used to perform a linear mapping transformation on the recurrence frequency data within the window, eliminating the inconsistency of dimensions caused by differences in production cycle time between different processes.

[0194] The normalized recurrence frequency index is weighted and fused with historical confidence data to generate a set of short-term memory factor values ​​that characterize causal stability.

[0195] By using sliding time window statistics and range standardization, the results of the previous step are transformed into a set of dimensionless short-term memory factor values, thereby achieving the expected technical effects of quantitative assessment of causal relationship stability and effective isolation of incidental noise.

[0196] S6.2: Receive the short-term memory factor value set, and use a preset occasional interference judgment threshold to traverse and compare the short-term memory factor value set one by one, and filter out a list of low confidence edge identifiers that are less than the occasional interference judgment threshold, so as to lock the occasional interference edge connection objects that need to be weakened.

[0197] The system receives a normalized set of short-term memory factor (STM) values, which contains stability metrics for each edge in the dynamic causal graph chain. It iterates through each element in the STM value set, extracting the unique identifier of the current edge to be evaluated and its corresponding STM value. A preset random interference threshold parameter is applied, set based on the statistical distribution boundary between noisy edges and true causal edges in historical production data, used to distinguish stable causal relationships from random fluctuations. The extracted STM value is compared with the random interference threshold to determine if the stability of the current edge is below the system tolerance limit. If the STM value is less than the random interference threshold, the edge is identified as a low-confidence random interference edge, and its unique identifier is added to the low-confidence edge identifier list. If the STM value is greater than or equal to the random interference threshold, the edge is identified as a high-confidence stable causal edge, the current iteration is skipped, and the next edge is processed. This comparison and filtering process is repeated until all elements in the STM value set have been iterated. By using threshold comparison and filtering, the set of short-term memory factor values ​​from the previous step is transformed into a list of low-confidence edge identifiers, which enables precise locking of occasional interfering connections in the dynamic causal graph. This provides a clear target for subsequent topological weight decay operations, effectively improving the purity and reliability of the quality attribution path.

[0198] For example, the threshold for determining occasional interference is set to 0.45. A set of short-term memory factor values ​​containing 5 edges is received: edge E1 is 0.82, edge E2 is 0.31, edge E3 is 0.67, edge E4 is 0.19, and edge E5 is 0.55. The set is traversed. Edge E1' (0.82) is greater than 0.45 and is determined to be a stable edge; edge E2' (0.31) is less than 0.45, and its identifier is added to the low-confidence edge identifier list; edge E3' (0.67) is greater than 0.45 and is determined to be a stable edge; edge E4' (0.19) is less than 0.45, and its identifier is added to the low-confidence edge identifier list; edge E5' (0.55) is greater than 0.45 and is determined to be a stable edge. The final output low-confidence edge identifier list contains [E2, E4]. This process accurately identifies occasional connections with low recurrence frequency and poor confidence, significantly improving the anti-interference capability of subsequent attribution analysis.

[0199] S6.3: For the occasional interference edge connection objects in the low confidence edge identifier list, perform a topology weight decay operation to reduce their connection strength in the dynamic causal graph link, and generate a pure causal subgraph structure after noise suppression, thereby eliminating the misleading effect of occasional interference on the attribution path.

[0200] Receive a list of low-confidence edge identifiers and the corresponding dynamic causal graph link data, and extract the current short-term memory factor value and historical topological weight value of each occasional interference edge connected object in the list.

[0201] Construct a weight update model based on the exponential decay mechanism. For each occasional interference edge in the list, read the process semantic fingerprint hash value of the source node and target node currently connected to it.

[0202] The weight decay coefficient is calculated. This coefficient is determined by the product of the preset base decay rate and the inverse of the current short-term memory factor. It is used to quantify the negative impact of noise interference on the stability of the causal chain.

[0203] The topology weight decay calculation is performed using the following formula: ; Among them, W new W represents the new topological weights after decay. old The original topological weights before decay, α being the basic decay rate constant, and M... f This represents the short-term memory factor value of the current edge.

[0204] The calculated new topological weight W new Replace the weight attribute of the corresponding edge in the original dynamic causal graph link to achieve numerical suppression of the connection strength of the occasional interference edge.

[0205] Traverse all edges in the dynamic causal graph link, identify and remove edge connections with topological weights below the minimum connectivity threshold, and cut off false causal paths formed by occasional interference.

[0206] The remaining high-weight edges and their associated process semantic fingerprint nodes are retained, the adjacency matrix of the local subgraph is reconstructed, and a pure causal subgraph structure after noise suppression is generated.

[0207] By using exponential decay weight updates and low-threshold pruning, the occasional interference edges locked in the previous step are transformed into weakened or broken ineffective connections, thereby achieving the expected technical effect of eliminating the misleading effect of occasional interference on the attribution path.

[0208] For example, setting the base attenuation rate α to 0.5, the original topological weight W of a certain occasional interference edge... old Its short-term memory factor M is 0.8. f The value is 0.2. Substituting this into the formula, we obtain the new weight W. new This equals 0.8 multiplied by e to the power of -2.5, resulting in approximately 0.065. If the minimum connectivity threshold is set to 0.1, then the edge weight of 0.065 is below the threshold, and the system directly removes this edge connection during the construction of the pure causal subgraph. For the other stable edge, its M... f The initial value was 0.9, and the calculated new weight was approximately 0.46, which was higher than the threshold and was retained. The final output pure causal subgraph only contained high-confidence causal links, significantly improving the accuracy of attribution paths.

[0209] S6.4: Based on the pure causal subgraph structure, extract the retained high-confidence process semantic fingerprint nodes and their corresponding dynamic confidence scores, and reassemble them linearly in series according to the process flow sequence to construct an initial process semantic chain composed of process semantic fingerprint nodes and dynamic confidence scores.

[0210] Receive a clean causal subgraph structure after noise suppression processing, which contains high-confidence procedural semantic fingerprint nodes and their associated edge attributes retained after short-term memory factor filtering.

[0211] Traverse all nodes in the pure causal subgraph, extract the process semantic fingerprint identifier corresponding to each node and the dynamic confidence score value calculated in step S5, and construct an initial data set containing node ID and confidence key-value pairs.

[0212] Obtain the standard process flow sequence definition table for the air conditioner rooftop integrated unit production line. This table clarifies the physical sequence and logical dependencies of all workstations from raw material loading to finished unit unloading.

[0213] Map and match the node IDs in the initial dataset with the standard process flow timing definition table, and assign an absolute timing position index on the actual production line to each retained process semantic fingerprint node.

[0214] Based on the allocated absolute temporal position index, the initial data set is sorted in ascending order to ensure that the node order strictly follows the physical flow direction of materials on the production line, eliminating logical jumps or reverse order phenomena caused by the graph traversal algorithm.

[0215] The sorted node sequence is linearly concatenated and recombined to generate a structured data object consisting of an ordered list of process semantic fingerprint nodes and their corresponding dynamic confidence scores, i.e., the initial process semantic chain.

[0216] By using the above-mentioned linear recombination processing method based on physical time constraints, the topological subgraph results obtained in the previous step are transformed into linear attribution chain data that conforms to the logic of the production process, thereby achieving the expected technical effect of readability and logical coherence of the quality problem tracing path.

[0217] For example, the pure causal subgraph contains three nodes: node A (condenser assembly, fingerprint hash: 0x1A2B, confidence level 0.92), node B (pipe welding, fingerprint hash: 0x3C4D, confidence level 0.88), and node C (air tightness test, fingerprint hash: 0x5E6F, confidence level 0.95). In the standard process timing definition table, the condenser assembly station number is 10, the pipe welding station number is 15, and the air tightness test station number is 20. The system maps nodes A, B, and C to timing indices 10, 15, and 20, respectively. After performing ascending sorting, the node order is confirmed as A→B→C. The final output initial process semantic chain is: [{"node_id":"0x1A2B", "process_name": "condenser assembly", "confidence": 0.92, "sequence_idx":10}, {"node_id": "0x3C4D", "process_name": "pipe welding", "confidence": 0.88, "sequence_idx": 15}, {"node_id": "0x5E6F", "process_name": "air tightness test", "confidence": 0.95, "sequence_idx": 20}]. This structure clearly shows the path of anomaly propagation along the production flow, significantly improving the accuracy of subsequent interpretable annotation generation.

[0218] S6.5: For each node pair in the initial process semantic chain, the process knowledge base is called to map the physical meaning of their shared discrete behavioral characteristics, generate interpretable annotation text containing mirror coupling descriptions or propagation attenuation coefficients, and finally output the final quality problem attribution path containing interpretable annotations.

[0219] Receive the initial process semantic chain, which contains a high-confidence process node sequence after noise suppression and its dynamic confidence score. Traverse adjacent node pairs in the initial process semantic chain and extract the shared discrete behavior feature vector between each pair of nodes calculated in step S5. The vector contains key indicators such as the probability of synchronization and the propagation attenuation coefficient.

[0220] A pre-built process knowledge base is invoked, which stores semantic mapping tables of standard process parameters, equipment physical characteristics, and common fault modes for each workstation of the integrated air conditioning unit. Using feature identifiers from the shared discrete behavior feature vectors as search keys, an exact matching query is performed in the process knowledge base to obtain the physical meaning description template of the corresponding feature.

[0221] For the retrieved physical meaning description template, the specific values ​​from the shared discrete behavioral feature vector are substituted into the variable placeholders in the template. For features involving mirror coupling, the correlation direction and strength of upstream and downstream process parameter fluctuations are calculated; for features involving propagation attenuation, the energy loss ratio of the signal transmitted from the upstream process to the downstream process is calculated.

[0222] Using natural language generation rules, the physical meaning description after filling in the numerical values ​​is structurally concatenated with the dynamic confidence score. A causal path description containing 'Process A-fingerprint ID' pointing to 'Process B-fingerprint ID' is generated, along with interpretable annotation text such as 'Welding current fluctuation mode and fastening torque attenuation are mirror-coupled, confidence level 0.86'.

[0223] The annotation text of all node pairs is concatenated according to the process flow sequence to form a complete attribution path report. Through process semantic mapping and natural language generation processing, the abstract fingerprint association is transformed into a quality problem attribution path with physical meaning, significantly improving the readability and verifiability of the attribution results.

[0224] For example, the initial process semantic chain includes nodes [Screw Tightening - F12] and [Pipe Welding - W07]. The shared feature vector shows a synchronization probability of 0.92 and a propagation attenuation coefficient of 0.15. The process knowledge base retrieves a 'mechanical stress transmission' mapping relationship between the torque anomaly of F12 and the current fluctuation of W07. The system substitutes the values ​​into the template and generates an annotation: 'The shortened steady-state duration of the F12 torque leads to unstable welding contact resistance of W07, which in turn causes high-frequency current fluctuations. The two are strongly mirror-coupled, with a dynamic confidence score of 0.86.' The final output includes a complete attribution path containing this annotation. Maintenance personnel can directly locate the pneumatic wrench pressure setting problem at the tightening station based on the annotation, significantly reducing troubleshooting time.

[0225] This application also provides a quality traceability device for an air conditioner rooftop unit production line, comprising: The barcode scanner is installed in each process unit of the production line to generate a unique traceability code for each air conditioner top-mounted unit. The industrial control terminal is connected to the barcode scanner and is used to execute the above-mentioned quality traceability method for the air conditioner top-mounted integrated unit production line, collect the process parameters generated by each process unit, and associate the process parameters generated by each process unit with the traceability code in real time. The traceability device is connected to the industrial control terminal and is used to retrieve, query, and display the status of each process.

[0226] By standardizing and integrating multi-source heterogeneous data from various process units using unified semantic anchoring, this invention solves the problems of traditional data fragmentation and the need for manual integration and comparison. Relying on process behavior pattern signatures, semantic fingerprint databases, and dynamic causal graph evolution mechanisms, it abandons the traditional full-domain data backtracking and static threshold attribution modes. This significantly reduces data processing resource overhead, improves the response speed of anomaly tracing and the efficiency of quality problem closure, and can adapt to complex production conditions such as small batches, multiple batches, and dynamic adjustment of process parameters on production lines, exhibiting excellent adaptability and traceability stability. At the same time, this invention breaks through the limitations of traditional methods that rely solely on numerical thresholds to determine anomalies. Through semantic deviation calculation and discrete behavioral feature analysis, it deeply mines hidden influencing factors such as operational behavior, process state transitions, and multi-parameter coupling, achieving fine-grained, interpretable, and accurate positioning of quality anomalies. This effectively reduces missed and false judgments, enabling rapid and accurate investigation of production anomalies, avoiding the production of batches of defective products, ensuring the stable and efficient operation of the production line, and significantly improving the intelligent quality control level and overall production efficiency of complex equipment production lines.

Claims

1. A quality traceability method for a production line of integrated rooftop air conditioner units, characterized in that, include: Acquire multi-source heterogeneous data from each process unit of the air conditioner rooftop unit production line, perform unified semantic anchoring processing on the multi-source heterogeneous data, and generate an original behavior pattern signature set. Based on the original behavioral pattern signature set, a hash vector representation of each process unit is generated, a process semantic fingerprint database is constructed, and the original behavioral pattern signature set is transformed into a process semantic fingerprint sequence. The cross-process influence relationship is statistically verified using the process semantic fingerprint sequence. The discrete process semantic fingerprint sequence is mapped to generate a basic causal skeleton graph structure with nodes as process semantic fingerprints and edges as lagging correlations. In response to the quality anomaly signal detected in real time, the current abnormal process fingerprint is used as the query starting point. A restricted depth-first traversal is performed on the basic causal skeleton graph and adjacent candidate process semantic fingerprints are dynamically loaded. The semantic deviation between the current abnormal process fingerprint and the candidate process semantic fingerprints is calculated, and a dynamically activated edge set is generated. Based on the discrete behavioral features shared in the set of dynamically activated edges, local causal strength re-evaluation is performed to generate a dynamic causal graph with short-term memory factors, which is then transformed into dynamic causal graph links. Based on the short-term memory factor threshold judgment result of the dynamic causal graph link, the occasional interference edge connection below the set threshold is weakened and the process semantic chain composed of process semantic fingerprint nodes and dynamic confidence scores is output, and the dynamic causal graph link is parsed into the final quality problem attribution path.

2. The quality traceability method for an air conditioner rooftop unit production line according to claim 1, characterized in that, The multi-source heterogeneous data includes structured sensor data, semi-structured equipment logs, and unstructured operation records for each key workstation; the original behavior pattern signature set includes the operation sequence topology, state transition frequency distribution, and abnormal event context association strength.

3. The quality traceability method for an air conditioner rooftop unit production line according to claim 1, characterized in that, The response to the quality anomaly signal detected in real time uses the current abnormal process fingerprint as the query starting point, performs a restricted depth-first traversal on the basic causal skeleton graph and dynamically loads the semantic fingerprints of adjacent candidate processes, calculates the semantic deviation between the current abnormal process fingerprint and the semantic fingerprints of candidate processes, and generates a dynamically activated edge set, specifically including: Acquire the quality anomaly signals detected in real time and the corresponding abnormal batch identifiers, extract the current abnormal process fingerprint from the process semantic fingerprint database based on the abnormal batch identifiers, and set the abnormal process fingerprint as the initial query starting node of the restricted depth-first traversal. Based on the topological connection relationship of the initial query starting node in the basic causal skeleton graph, the upstream adjacent process node directly connected to the initial query starting node is located, and the latest batch process semantic fingerprint sequence corresponding to the upstream adjacent process node is dynamically loaded from the process semantic fingerprint database according to the preset recent batch number threshold. The abnormal process fingerprints contained in the initial query starting node are mapped to each candidate process semantic fingerprint in the latest batch process semantic fingerprint sequence via vector space mapping. The discrete feature difference values ​​between the abnormal process fingerprints and each candidate process semantic fingerprint are calculated to generate the original semantic deviation value set. Based on the comparison and judgment between the original set of semantic deviation values ​​and the preset semantic deviation threshold, abnormal associated nodes with a value greater than the semantic deviation threshold are selected, and the topological connection between the abnormal associated nodes and the initial query starting node is marked as a high-confidence potential causal edge to generate a set of edges to be verified. Based on the set of edges to be verified, a recursive restricted depth-first traversal operation is performed. The latest abnormal associated node in the traversal process is updated to the new current query starting node, and the adjacent node loading and deviation calculation process is repeated until the traversal depth reaches the maximum level limit or no new high-confidence potential causal edges are generated. Finally, a dynamic active edge set consisting of all marked high-confidence potential causal edges is output.

4. A quality traceability method for an air conditioner rooftop unit production line according to claim 3, characterized in that, The preset threshold for the number of most recent batches is dynamically adjusted based on the stability of the production line cycle time. When the production line cycle time is stable and the consistency of each process operation is high, the threshold for the number of most recent batches is 3-5; when the production line cycle time is unstable and the process operation fluctuates greatly, the threshold for the number of most recent batches is 10-15.

5. A quality traceability method for an air conditioner rooftop unit production line according to claim 3, characterized in that, The discrete feature difference between the abnormal process fingerprint and the semantic fingerprint of each candidate process is calculated using the Hamming distance algorithm.

6. A quality traceability method for an air conditioner rooftop unit production line according to claim 1, characterized in that, The process of re-evaluating local causal strength and generating a dynamic causal graph with a short-term memory factor based on the shared discrete behavioral features in the set of dynamically activated edges, and transforming it into a dynamic causal graph link, specifically includes: Discrete behavioral feature intersection operations are performed on the semantic fingerprint nodes of upstream and downstream processes in the dynamically activated edge set to extract shared discrete behavioral feature vectors, including the probability of synchronous occurrence and the propagation attenuation coefficient, in order to construct a feature input benchmark set for local causal strength re-evaluation. Based on the shared discrete behavior feature vector, the conditional probability of the initial topological connection of each edge in the dynamically activated edge set is calculated to generate a real-time posterior confidence value that characterizes the credibility of causal association under the current abnormal batch. Based on the real-time posterior confidence value, the historical abnormal event recurrence frequency statistics module is called to obtain the recurrence frequency data of the corresponding edge in the last M abnormal events, and the recurrence frequency data and the real-time posterior confidence value are weighted and fused to generate a comprehensive causal strength index including timeliness weight. Based on the comprehensive causal strength index, a short-term memory factor mapping function operation is performed to convert the comprehensive causal strength index into a short-term memory factor value that characterizes the stability of the causal chain. The edge attributes in the dynamically activated edge set are enhanced and labeled using the short-term memory factor values, and the edge connections with short-term memory factor values ​​are recombined into dynamic causal graph links with recurrence frequency weighting attributes.

7. A quality traceability method for an air conditioner rooftop unit production line according to claim 1, characterized in that, The step of determining the short-term memory factor threshold based on the dynamic causal graph link, weakening occasional interference edges below a set threshold, and outputting a process semantic chain composed of process semantic fingerprint nodes and dynamic confidence scores, and parsing the dynamic causal graph link into the final quality problem attribution path, specifically includes: The recurrence frequency weighted attribute and historical confidence data of each edge in the dynamic causal graph link are obtained. The recurrence frequency weighted attribute is normalized based on the sliding time window statistical mechanism to generate a set of short-term memory factor values ​​that characterize causal stability. Receive the set of short-term memory factor values, and use a preset occasional interference judgment threshold to traverse and compare the set of short-term memory factor values ​​one by one, and filter out a list of low confidence edge identifiers that are less than the occasional interference judgment threshold. For the occasional interfering edge connection objects in the low confidence edge identifier list, a topology weight decay operation is performed to reduce their connection strength in the dynamic causal graph link, generating a clean causal subgraph structure after noise suppression. Based on the pure causal subgraph structure, the high-confidence process semantic fingerprint nodes and their corresponding dynamic confidence scores are extracted and retained, and linearly reassembled according to the process flow sequence to construct an initial process semantic chain composed of process semantic fingerprint nodes and dynamic confidence scores. For each node pair in the initial process semantic chain, the process knowledge base is invoked to map the physical meaning of their shared discrete behavioral features, generate interpretable annotation text, and output the final quality problem attribution path containing the interpretable annotation text.

8. A quality traceability method for a rooftop air conditioner integrated unit production line according to claim 7, characterized in that, The interpretable annotation text contains a mirrored coupling description or a propagation attenuation coefficient.

9. A quality traceability method for an air conditioner rooftop unit production line according to claim 7, characterized in that, The preset threshold parameter for determining occasional interference is set based on the statistical distribution boundary between noise edges and true causal edges in historical production data.

10. A quality traceability device for a production line of integrated rooftop air conditioner units, characterized in that, include: The barcode scanner is installed in each process unit of the production line to generate a unique traceability code for each air conditioner top-mounted unit. An industrial control terminal, whose signal is connected to the barcode scanner, is used to execute the quality traceability method for the air conditioner top-mounted integrated unit production line as described in any one of claims 1-9, collect the process parameters generated by each process unit, and associate the process parameters generated by each process unit with the traceability code in real time. The traceability device is connected to the industrial control terminal and is used to retrieve, query, and display the status of each process.

Images