Big data-based digital management system for automobile generator assembly production line

Abstract

This invention discloses a big data-based digital management system for automotive generator assembly lines, belonging to the field of intelligent manufacturing control technology. It includes a system data preprocessing module, a state decoding and analysis module, an assembly deviation analysis module, a correlation graph construction module, and an anomaly pattern mining module. The system performs time-series alignment and anomaly marking on sensor sequence data, and state encoding and change point detection on equipment state parameter data. A state transition matrix is ​​constructed based on the topological constraints of the assembly process, and a dual-datastream joint state sequence decoding is completed using an improved Hidden Markov Model. An assembly deviation mark set is generated by comparing the actual operating state of the workstation with standard process segments. A deviation-material correlation graph is constructed by combining material traceability data. Cross-workstation anomaly correlation patterns are extracted through frequent subgraph mining and pushed to the digital management interface, adapting to the refined control requirements of automotive generator assembly lines.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing control technology, specifically a digital management system for automotive generator assembly production lines based on big data. Background Technology

[0002] Automotive generator assembly lines operate in a multi-station continuous operation scenario, continuously generating sensor sequence data, assembly equipment status parameter data, and component material traceability data during production. Existing digital management systems mostly adopt a unified and universal data processing logic, failing to establish differentiated processing mechanisms for different types of raw production data. Sensor time-series data suffers from timing errors and the mixing of invalid data. Equipment status parameters lack standardized coding and abrupt change feature identification processes. The raw data cannot form a standardized data stream with a unified structure and matching timing, making it difficult to meet the data foundation requirements for refined analysis of assembly processes.

[0003] Conventional production line status analysis methods often employ a single-data-dimension independent analysis mode. The analysis models used do not incorporate the actual topological constraints of the assembly process for parameter construction, making it impossible to achieve collaborative decoding of sensor data streams and equipment status streams. The determination of the assembly process's operating status only involves local comparisons at a single workstation, lacking a mechanism for linking assembly deviations with component material traceability information, and failing to analyze the inherent connections between deviations at each workstation from a global perspective.

[0004] Existing technologies cannot deeply mine the hidden abnormal correlation features that repeatedly exist between multiple workstations. Hidden problems in the assembly process of the production line remain for a long time, making it difficult to achieve a global analysis of the assembly operation status and a systematic sorting of abnormal features. This cannot support the actual application needs of digital and refined management and control of automotive generator assembly production lines. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; To this end, the present invention proposes a digital management system for automotive generator assembly production lines based on big data, comprising: The data preprocessing module acquires raw data including sensor sequence data, assembly equipment status parameter data, and component material traceability data. It performs time alignment and anomaly marking on the sensor sequence data to generate a regularized sensor data stream. It also performs status coding and change point detection on the assembly equipment status parameter data to generate a regularized equipment status stream. The state decoding analysis module uses an improved hidden Markov model to perform joint state sequence decoding processing on the regularized sensor data stream and the regularized equipment state stream. The improved hidden Markov model constructs a state transition matrix based on the topological constraints of the assembly process. The assembly deviation analysis module identifies the actual operating state segments of each assembly station in the automotive generator assembly production line based on the results of the joint state sequence decoding process, and compares the actual operating state segments with the preset standard assembly process segments to generate an assembly deviation mark set. The correlation graph construction module constructs a deviation-material correlation graph for each assembly station based on the assembly deviation mark set and the component material traceability data. The anomaly pattern mining module performs frequent subgraph mining processing on the deviation-material association graph, extracts the abnormal association patterns that recur across assembly stations, and pushes them to the digital management interface of the automotive generator assembly production line.

[0006] Furthermore, the improved Hidden Markov Model constructs a state transition matrix based on the topological constraints of the assembly process, including: Obtain the standard assembly process sequence of the automotive generator assembly production line. The standard assembly process sequence includes multiple assembly process nodes arranged in the order of assembly and the process dependencies between each assembly process node. Each assembly process node is mapped to a hidden state of the improved Hidden Markov Model, and the process dependencies are mapped to allowable transition direction constraints between hidden states, wherein the previous process node can only transfer to the next process node or remain at the current process node, and transfer from the next process node to the previous process node is prohibited. An initial state transition matrix is ​​constructed based on the allowed transition direction constraints. The matrix elements corresponding to the allowed transition directions in the initial state transition matrix are set as variable weight parameters to be trained, and the matrix elements corresponding to the prohibited transition directions in the initial state transition matrix are fixed to zero values. The parallel process merging constraint contained in the assembly process topology constraint is incorporated into the initial state transition matrix. The parallel process merging constraint is used to indicate that multiple parallel process nodes are allowed to transition to each other during execution and have the same transition probability. The initial state transition matrix after being processed by the allowed transition direction constraint and the parallel process merging constraint is subjected to row normalization to generate a state transition matrix that satisfies the topological constraints of the assembly process.

[0007] Furthermore, time alignment and anomaly marking are performed on the sensor sequence data to generate a regularized sensor data stream, including: Sensor sequence data for each data acquisition node is extracted from the raw data. The sensor sequence data includes torque sensor numerical sequences, displacement sensor numerical sequences, and vibration sensor numerical sequences. The timestamp fields corresponding to the torque sensor numerical sequence, the displacement sensor numerical sequence, and the vibration sensor numerical sequence are identified. The three sequences are resampled according to the same time granularity using the timestamp fields as the alignment reference to obtain a time-aligned multi-channel sensor matrix. Perform sliding window statistical processing on each channel in the time-aligned multi-channel sensor matrix to calculate the local mean and local standard deviation of the sensor values ​​within each sliding window; Each sensor value within each sliding window is compared with the local mean and the local standard deviation. When a sensor value exceeds the range of the local mean plus or minus three times the local standard deviation, the sensor value is marked as an outlier, and the outlier value is replaced with the local median value to generate a regularized sensor data stream.

[0008] Furthermore, the step of performing state coding and change point detection processing on the assembly equipment state parameter data to generate a regularized equipment state flow includes: The assembly equipment status parameter data for each assembly equipment is extracted from the raw data. The assembly equipment status parameter data includes a sequence of equipment operating current parameters, a sequence of equipment operating temperature parameters, and a sequence of equipment operating speed parameters. Each current value in the equipment operating current parameter sequence is mapped to a discrete current state code according to a preset current state partition table; each temperature value in the equipment operating temperature parameter sequence is mapped to a discrete temperature state code according to a preset temperature state partition table; and each speed value in the equipment operating speed parameter sequence is mapped to a discrete speed state code according to a preset speed state partition table. The discrete current state code, the discrete temperature state code, and the discrete speed state code at the same time point are concatenated to generate a combined state code sequence corresponding to the time point. The combined state code sequence is subjected to cumulative sum and change point detection processing. The cumulative sum statistic of the combined state code at each time point is calculated. When the cumulative sum statistic of multiple consecutive time points exceeds the preset change threshold, the time point is marked as a device state change point. The combined state code sequence between two adjacent device state change points is taken as a steady-state operation segment. The combined state code that appears most frequently in each steady-state operation segment is taken as the representative state code of the steady-state operation segment. The representative state codes of all steady-state operation segments are arranged in chronological order to generate a regular device state flow.

[0009] Furthermore, the joint state sequence decoding processing of the regularized sensor data stream and the regularized device state stream using an improved Hidden Markov Model includes: The regularized sensor data stream is used as the observation sequence input of the improved hidden Markov model, and the regularized device state stream is used as the auxiliary observation sequence input of the improved hidden Markov model. Construct the emission probability matrix of the improved hidden Markov model, where each element in the emission probability matrix represents the joint generation probability of the observation sequence and the auxiliary observation sequence in a given hidden state; Each observation vector in the observation sequence is concatenated with each auxiliary observation vector in the auxiliary observation sequence to generate a joint observation vector sequence; Based on the emission probability matrix and the state transition matrix that satisfies the topological constraints of the assembly process, Viterbi decoding is performed on the joint observation vector sequence to solve for the hidden state sequence that maximizes the generation probability of the joint observation vector sequence. The hidden state sequence is output as the result of decoding the joint state sequence, and each hidden state in the hidden state sequence corresponds to the assembly process node of the automobile generator assembly line at the corresponding time point.

[0010] Furthermore, based on the results of the joint state sequence decoding process, the actual operating state segments of each assembly station in the automotive generator assembly production line are identified, and these actual operating state segments are compared with preset standard assembly process segments to generate an assembly deviation mark set, including: Based on the hidden state corresponding to each time point in the result of the joint state sequence decoding process, the hidden state is mapped to the assembly process node name in the automobile generator assembly production line. Traverse the hidden state sequence in chronological order, extract the time intervals where multiple consecutive time points and hidden states remain unchanged as an actual running state segment, and each actual running state segment contains a start timestamp, an end timestamp and the corresponding assembly process node name. Obtain a set of standard assembly process segments for the automotive generator assembly production line. The set of standard assembly process segments includes a standard start timestamp, a standard end timestamp, and a standard duration range for each assembly process node. The actual running state segment is compared with the standard assembly process segment with the same assembly process node name in the standard assembly process segment set, and the difference between the actual duration interval of the actual running state segment and the standard duration interval is calculated. When the absolute value of the difference exceeds the preset duration deviation threshold, a duration deviation mark corresponding to the assembly process node is generated, and the duration deviation mark is added to the assembly deviation mark set.

[0011] Furthermore, based on the assembly deviation mark set and the component material traceability data, a deviation-material correlation graph is constructed for each assembly station, including: The deviation-material correlation diagram is used to describe the mapping relationship between assembly deviation types and batches of parts used; Extract component material traceability data from the raw data. The component material traceability data includes the component batch number used in each assembly cycle at each assembly station and the supplier information corresponding to each component batch number. Each duration deviation mark in the assembly deviation mark set is associated with the assembly process node that generates the duration deviation mark, and then the corresponding assembly station and the corresponding assembly cycle are located through the assembly process node. Obtain the batch number of the parts used by the assembly station during the assembly cycle, use the batch number of the parts as a material node in the deviation-material association diagram, and use the duration deviation mark as a deviation node in the deviation-material association diagram; Establish an association edge between the deviation node and the material node, and the weight of the association edge is equal to the frequency of the deviation node and the material node appearing together in the same assembly cycle; Aggregate all deviation nodes, all material nodes, and all associated edges on the same assembly station to generate a deviation-material association graph corresponding to the assembly station.

[0012] Furthermore, frequent subgraph mining is performed on the deviation-material association graph to extract abnormal association patterns that recur across assembly stations, including: All deviation-material association maps corresponding to assembly stations are merged into a single overall deviation-material association map, which includes all deviation nodes, all material nodes, and all associated edges. Set a subgraph support threshold, traverse all subgraph structures in the overall deviation-material association graph whose number of nodes is not greater than the preset maximum subgraph size, and calculate the frequency of each subgraph structure in the overall deviation-material association graph. Subgraph structures whose frequency is greater than or equal to the subgraph support threshold are selected as frequent subgraph candidate sets. For each frequent subgraph in the frequent subgraph candidate set, extract the deviation type corresponding to the deviation node and the component batch number corresponding to the material node contained in the frequent subgraph, and use the correspondence between the deviation type and the component batch number as the abnormal association mode. The abnormal association patterns are sorted from high to low according to the frequency of their corresponding frequent subgraphs in the overall deviation-material association diagram, and the abnormal association patterns with the highest ranking are extracted as abnormal association patterns that recur across assembly stations.

[0013] Furthermore, the system also includes The root cause tracing module is used to perform root cause tracing processing on the abnormal association patterns and generate a root cause tracing report, specifically including: Obtain the target component batch number in the abnormal association pattern that occurs repeatedly across assembly stations, and query the target supplier information and target production date information corresponding to the target component batch number from the component material traceability data; Extract all assembly deviation markers associated with the batch number of the target component from the original data, and group all assembly deviation markers according to assembly station and assembly time to obtain a deviation distribution matrix; Principal component analysis is performed on the deviation distribution matrix to reduce its dimensionality, and the principal component directions with the highest contribution rates are extracted. Assembly stations whose load coefficients on the principal component directions exceed a preset load threshold are marked as high-risk stations. Based on the target supplier information, the target production date information, and the location information of the high-risk workstation, a root cause tracing path is constructed according to the time sequence and workstation sequence. The root cause tracing path starts from the target component batch number, passes through the high-risk workstation, and arrives at the corresponding assembly deviation mark. The root cause tracing path is output as a root cause tracing report to the digital management interface of the automotive generator assembly line.

[0014] Further, based on the emission probability matrix and the state transition matrix satisfying the assembly process topology constraints, Viterbi decoding is performed on the joint observation vector sequence to solve for the hidden state sequence that maximizes the generation probability of the joint observation vector sequence, including: Initialize the Viterbi path probability matrix, input the first joint observation vector of the joint observation vector sequence into the emission probability matrix, calculate the probability value of each hidden state at the initial time, and set the predecessor state corresponding to each hidden state to empty; Starting from the second joint observation vector in the joint observation vector sequence, perform iterative calculation processing on each joint observation vector in the joint observation vector sequence. For each hidden state, calculate the product of the transition probability from all hidden states in the previous time step to the current hidden state and the corresponding path probability in the previous time step. The maximum value is selected from the product as the cumulative path probability of the current hidden state at the current time, and the hidden state from the previous time that generated the maximum value is recorded as the predecessor state of the current hidden state. The cumulative path probability and the predecessor state are recorded in the corresponding positions of the Viterbi path probability matrix. After the iterative calculation is completed, the hidden state with the largest cumulative path probability at the last time step in the Viterbi path probability matrix is ​​selected as the starting point of the final state. Starting from the initial point of the final state, based on the preceding state information recorded in the Viterbi path probability matrix, we backtrack to the initial time to obtain the hidden state sequence that maximizes the probability of generating the joint observation vector sequence.

[0015] Compared with the prior art, the beneficial effects of the present invention are: For sensor sequence data, time-series alignment and anomaly labeling are performed; for assembly equipment status parameter data, status coding and change point detection are performed. These two types of raw production data undergo independent processing and organization based on their own data attributes, resulting in heterogeneous data forming standardized data streams with unified format and temporal order. The disorganized information of different categories of raw data is organized and streamlined, preserving the data's inherent temporal logic and state change characteristics. Multi-source production data can adapt to the input specifications of subsequent model analysis, improving the usability and organization of raw production data and meeting the data adaptation requirements for full-process data analysis of the assembly process.

[0016] Based on the topological constraints of the assembly process, the state transition matrix of the improved Hidden Markov Model is constructed, enabling joint state sequence decoding of the regularized sensor data stream and the regularized equipment state stream. The two types of data streams are incorporated into the same analysis framework for collaborative analysis. The actual operating state segments of each assembly station are compared item by item with the standard assembly process segments to generate a complete set of assembly deviation markers. The deviation information is systematically collected and presented in a structured manner.

[0017] By combining assembly deviation marker sets and component material traceability data, a deviation-material correlation graph for each workstation is constructed, forming a structured relationship between assembly deviations and material information at each workstation. Frequent subgraph mining is performed on the deviation-material correlation graph, automatically filtering and extracting similar abnormal correlation features across assembly workstations. Discrete abnormal information scattered across various workstations forms an integrated correlation representation, and the correlational abnormal features hidden in the assembly process can be autonomously sorted and collected. Attached Figure Description

[0018] Figure 1 This is a sequence diagram of the big data-based digital management system for automotive generator assembly lines described in this invention. Figure 2 A flowchart for constructing the state transition matrix for an improved hidden Markov model; Figure 3 This is a flowchart for time-series alignment and anomaly marking processing of sensor sequence data. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1 A specific implementation of a big data-based digital management system for automotive generator assembly lines is as follows: This big data-based digital management system for automotive generator assembly lines includes a data preprocessing module, a state decoding and analysis module, an assembly deviation analysis module, a correlation graph construction module, and an anomaly pattern mining module. The data preprocessing module acquires raw data including sensor sequence data, assembly equipment state parameter data, and component material traceability data. It performs time-series alignment and anomaly marking processing on the sensor sequence data to generate a regularized sensor data stream. It performs state encoding and change point detection processing on the assembly equipment state parameter data to generate a regularized equipment state stream. The state decoding and analysis module uses an improved Hidden Markov Model (HMM) to perform joint state sequence decoding processing on the regularized sensor data stream and the regularized equipment state stream. The improved HMM constructs a state transition matrix based on the topological constraints of the assembly process. The assembly deviation analysis module identifies the actual operating state segments of each assembly station in the automotive generator assembly line based on the results of the joint state sequence decoding processing. It then compares the actual operating state segments with preset standard assembly process segments to generate an assembly deviation mark set. The correlation graph construction module constructs a deviation-material correlation graph for each assembly station based on the assembly deviation marker set and the component material traceability data. The anomaly pattern mining module performs frequent subgraph mining processing on the deviation-material correlation graph, extracts the abnormal correlation patterns that recur across assembly stations, and pushes them to the digital management interface of the automotive generator assembly production line.

[0021] In one embodiment of the present invention, the specific implementation of constructing the state transition matrix based on the topological constraints of the assembly process using the improved Hidden Markov Model is as follows: (See attached document). Figure 2The standard assembly process sequence of the automotive generator assembly production line is obtained. The standard assembly process sequence includes multiple assembly process nodes arranged in the assembly sequence and the process dependencies between each assembly process node. Each assembly process node is mapped to a hidden state of the improved Hidden Markov Model. The process dependencies are mapped to allowed transition direction constraints between hidden states. The preceding process node can only transition to the following process node or remain at the current process node, and transition from the following process node to the preceding process node is prohibited. An initial state transition matrix is ​​constructed based on the allowed transition direction constraints. The matrix elements corresponding to the allowed transition directions in the initial state transition matrix are set as variable weight parameters to be trained. The matrix elements corresponding to the prohibited transition directions in the initial state transition matrix are fixed to zero. The parallel process merging constraint included in the assembly process topology constraint is incorporated into the initial state transition matrix. The parallel process merging constraint is used to indicate that multiple parallel process nodes are allowed to transition to each other during execution and have the same transition probability. Row normalization is performed on the initial state transition matrix after processing by the allowed transition direction constraints and the parallel process merging constraints to generate a state transition matrix that satisfies the assembly process topology constraints.

[0022] In a specific implementation, taking an automotive generator assembly line as an example, the process of constructing a state transition matrix based on the topological constraints of the assembly process using the improved Hidden Markov Model is as follows: Obtain the standard assembly process sequence of the automotive generator assembly line. The standard assembly process sequence includes multiple assembly process nodes arranged in the order of assembly and the process dependencies between each assembly process node. The assembly process nodes include stator assembly process nodes, rotor assembly process nodes, bearing press-fit process nodes, end cover assembly process nodes, and performance testing process nodes. The process dependencies indicate that the next process node can only be executed after the previous process node is completed.

[0023] In some embodiments, each assembly process node is mapped to a hidden state of the improved Hidden Markov Model, and the process dependencies are mapped to allowable transition direction constraints between hidden states, wherein the previous process node can only transition to the next process node or remain at the current process node, and transition from the next process node to the previous process node is prohibited; an initial state transition matrix is ​​constructed based on the allowable transition direction constraints, and the matrix elements corresponding to the allowed transition directions in the initial state transition matrix are set as variable weight parameters to be trained, while the matrix elements corresponding to the prohibited transition directions in the initial state transition matrix are fixed to zero values.

[0024] Optionally, the parallel process merging constraint included in the assembly process topology constraint is incorporated into the initial state transition matrix. The parallel process merging constraint is used to indicate that multiple parallel process nodes are allowed to transition to each other during execution with the same transition probability. In the automotive generator assembly production line, the parallel process merging constraint indicates that the stator assembly process node and the rotor assembly process node are allowed to transition to each other during execution with the same transition probability.

[0025] In some embodiments, row normalization is performed on the initial state transition matrix after processing with the allowed transition direction constraint and the parallel process merging constraint to generate a state transition matrix that satisfies the topological constraints of the assembly process; the formula for the row normalization is expressed as: in: This represents the value of the element in the i-th row and j-th column of the state transition matrix after row normalization. This represents the variable weight parameter or zero value in the i-th row and j-th column of the initial state transition matrix before row normalization, N represents the total number of hidden states in the improved Hidden Markov Model, and k is the summation index variable. This represents the sum of all elements in the i-th row of the initial state transition matrix. It can be understood that the sum of the elements in each row of the state transition matrix is ​​1, where the matrix elements corresponding to allowed transition directions have non-zero probability values ​​after row normalization, and the matrix elements corresponding to prohibited transition directions remain zero after row normalization. It can be understood that the improved Hidden Markov Model, in subsequent decoding, only allows hidden states to transition according to the allowed transition direction constraints and the parallel process merging constraints, and prohibits any state jumps that violate the assembly process topology constraints.

[0026] In one embodiment of the present invention, the specific implementation method for performing time alignment and anomaly marking processing on sensor sequence data to generate a regularized sensor data stream is as follows, see [reference]. Figure 3The sensor sequence data of each data acquisition node is extracted from the raw data. The sensor sequence data includes torque sensor value sequences, displacement sensor value sequences, and vibration sensor value sequences. The timestamp fields corresponding to the torque sensor value sequences, displacement sensor value sequences, and vibration sensor value sequences are identified. Using the timestamp fields as an alignment reference, the three sequences are resampled at the same time granularity to obtain a time-aligned multi-channel sensor matrix. Sliding window statistical processing is performed on each channel in the time-aligned multi-channel sensor matrix to calculate the local mean and local standard deviation of the sensor values ​​within each sliding window. Each sensor value within each sliding window is compared with the local mean and the local standard deviation. When the sensor value exceeds the range of the local mean plus or minus three times the local standard deviation, the sensor value is marked as an outlier and the value of the outlier is replaced with the local median value to generate a regularized sensor data stream.

[0027] The specific implementation method for generating a regularized equipment state flow by performing state coding and change point detection processing on the assembly equipment state parameter data is as follows: The assembly equipment state parameter data for each assembly piece of equipment is extracted from the original data. This assembly equipment state parameter data includes a sequence of equipment operating current parameters, a sequence of equipment operating temperature parameters, and a sequence of equipment operating speed parameters. Each current value in the equipment operating current parameter sequence is mapped to a discrete current state code according to a preset current state partitioning table; each temperature value in the equipment operating temperature parameter sequence is mapped to a discrete temperature state code according to a preset temperature state partitioning table; and each speed value in the equipment operating speed parameter sequence is mapped to a discrete speed state code according to a preset speed state partitioning table. Simultaneously... The discrete current state code, discrete temperature state code, and discrete speed state code at each time point are concatenated to generate a combined state code sequence corresponding to that time point. Cumulative sum and change point detection processing is performed on the combined state code sequence to calculate the cumulative sum statistic of the combined state code at each time point. When the cumulative sum statistic of multiple consecutive time points exceeds a preset change threshold, the time point is marked as a device state change point. The combined state code sequence between two adjacent device state change points is taken as a steady-state operation segment. The combined state code that appears most frequently in each steady-state operation segment is taken as the representative state code of the steady-state operation segment. The representative state codes of all steady-state operation segments are arranged in chronological order to generate a regularized device state flow.

[0028] In practical implementation, taking an assembly station on an automotive generator assembly line as an example, the process of performing time alignment and outlier marking on sensor sequence data to generate a regularized sensor data stream is as follows: Extract torque sensor value sequences, displacement sensor value sequences, and vibration sensor value sequences from the original data; identify the timestamp field corresponding to each sequence; resample the three sequences according to the same time granularity using the timestamp field as the alignment benchmark to obtain a time-aligned multi-channel sensor matrix; perform sliding window statistical processing on each channel in the multi-channel sensor matrix, calculate the local mean and local standard deviation of the sensor values ​​within each sliding window; compare each sensor value within each sliding window with the range of the local mean plus or minus three times the local standard deviation; when a sensor value exceeds this range, mark the value as an outlier and replace it with the local median value to generate a regularized sensor data stream.

[0029] In some embodiments, the process of generating a regular equipment state flow by performing state coding and change point detection processing on the assembly equipment state parameter data is as follows: Extract the equipment operating current parameter sequence, equipment operating temperature parameter sequence, and equipment operating speed parameter sequence from the original data; map each current value to a discrete current state code according to a preset current state partition table; map each temperature value to a discrete temperature state code according to a preset temperature state partition table; map each speed value to a discrete speed state code according to a preset speed state partition table; and concatenate the three discrete state codes at the same time point to generate a combined state code sequence. Optionally, cumulative sum and change point detection processing is performed on the combined state code sequence. The cumulative sum statistic of the combined state code at each time point is calculated. When the cumulative sum statistic of multiple consecutive time points exceeds a preset change threshold, that time point is marked as a device state change point. The combined state code sequence between two adjacent device state change points is considered as a steady-state operating segment. The combined state code that appears most frequently in each steady-state operating segment is taken as the representative state code of that segment. The representative state codes of all steady-state operating segments are arranged in chronological order to generate a regularized device state flow. The formula for calculating the cumulative sum statistic is: in: This represents the cumulative sum statistic at time point t, where t is the index of the time point and is a positive integer. This represents the combined state code value at time point i. Let represent the mean of the combined state-encoded sequence during the initial stable phase, and ...

[0030] In one embodiment of the present invention, the specific implementation of joint state sequence decoding processing of the regularized sensor data stream and the regularized equipment state stream using an improved Hidden Markov Model is as follows: the regularized sensor data stream is used as the observation sequence input of the improved Hidden Markov Model, and the regularized equipment state stream is used as the auxiliary observation sequence input of the improved Hidden Markov Model. The emission probability matrix of the improved Hidden Markov Model is constructed. Each element in the emission probability matrix is ​​used to represent the joint generation probability of the observation sequence and the auxiliary observation sequence under a given hidden state. Each observation vector in the observation sequence is concatenated with each auxiliary observation vector in the auxiliary observation sequence to generate a joint observation vector sequence. Based on the emission probability matrix and the state transition matrix that satisfies the topological constraints of the assembly process, Viterbi decoding processing is performed on the joint observation vector sequence to solve for the hidden state sequence that maximizes the generation probability of the joint observation vector sequence. The hidden state sequence is output as the result of the joint state sequence decoding processing. Each hidden state in the hidden state sequence corresponds to the assembly process node of the automobile generator assembly line at the corresponding time point.

[0031] The specific implementation of Viterbi decoding processing to solve for the hidden state sequence that maximizes the generation probability of the joint observation vector sequence based on the emission probability matrix and the state transition matrix satisfying the assembly process topology constraints is as follows: Initialize the Viterbi path probability matrix, input the first joint observation vector of the joint observation vector sequence into the emission probability matrix, calculate the probability value of each hidden state at the initial time, and set the predecessor state corresponding to each hidden state to empty. Starting from the second joint observation vector of the joint observation vector sequence, perform iterative calculation processing on each joint observation vector in the joint observation vector sequence. For each hidden state, calculate the transition from all hidden states at the previous time step. The product of the transition probability of the current hidden state and the corresponding path probability at the previous time step is used to select the maximum value of the product as the cumulative path probability of the current hidden state at the current time step. The hidden state from the previous time step that generated the maximum value is recorded as the predecessor state of the current hidden state. The cumulative path probability and the predecessor state are recorded in the corresponding positions of the Viterbi path probability matrix. After the iterative calculation is completed, the hidden state with the largest cumulative path probability at the last time step in the Viterbi path probability matrix is ​​selected as the starting point of the final state. Starting from the starting point of the final state, the process is reversed back to the initial time step based on the predecessor state information recorded in the Viterbi path probability matrix to obtain the hidden state sequence that maximizes the generation probability of the joint observation vector sequence.

[0032] In a specific implementation, taking an automotive generator assembly line as an example, the process of using an improved Hidden Markov Model to perform joint state sequence decoding on the regularized sensor data stream and the regularized equipment state stream is as follows: The regularized sensor data stream is used as the observation sequence input of the improved Hidden Markov Model, and the regularized equipment state stream is used as the auxiliary observation sequence input of the improved Hidden Markov Model; the emission probability matrix of the improved Hidden Markov Model is constructed, where each element of the emission probability matrix represents the joint generation probability of the observation sequence and the auxiliary observation sequence under a given hidden state; each observation vector in the observation sequence is concatenated with each auxiliary observation vector in the auxiliary observation sequence to generate a joint observation vector sequence.

[0033] In some embodiments, based on the emission probability matrix and the state transition matrix that satisfies the topological constraints of the assembly process, Viterbi decoding is performed on the joint observation vector sequence to solve for the hidden state sequence that maximizes the generation probability of the joint observation vector sequence; the hidden state sequence is output as the result of the joint state sequence decoding process, and each hidden state in the hidden state sequence corresponds to the assembly process node of the automobile generator assembly line at the corresponding time point.

[0034] Optionally, the specific process of performing Viterbi decoding on the joint observation vector sequence is as follows: Initialize the Viterbi path probability matrix, input the first joint observation vector of the joint observation vector sequence into the emission probability matrix, calculate the probability value of each hidden state at the initial time, and set the predecessor state corresponding to each hidden state to empty; starting from the second joint observation vector of the joint observation vector sequence, perform iterative calculation processing on each joint observation vector, for each hidden state, calculate the product of the transition probability from all hidden states at the previous time to the current hidden state and the corresponding path probability at the previous time, select the maximum value from the product as the cumulative path probability of the current hidden state at the current time, and record the hidden state from the previous time that generates the maximum value as the predecessor state of the current hidden state, and record the cumulative path probability and the predecessor state in the corresponding positions of the Viterbi path probability matrix. In a specific implementation, the formula for calculating the cumulative path probability in the iterative calculation processing is: in: This represents the cumulative path probability of being in hidden state j at time t, where t is the time index and t≥2, and j is the index of the current hidden state. This represents the cumulative path probability of being in hidden state i at time t-1, where i is the index of the hidden state at the previous time step and its value ranges from 1 to the total number of hidden states. Let represent the transition probability from hidden state i to hidden state j. This indicates the generation of a joint observation vector in hidden state j. The probability of launch, This represents taking the maximum value for all possible values ​​of i. In some embodiments, after the iterative calculation is completed, the hidden state with the largest cumulative path probability at the last moment in the Viterbi path probability matrix is ​​selected as the starting point of the final state; starting from the starting point of the final state, the process is reversed back to the initial moment based on the predecessor state information recorded in the Viterbi path probability matrix to obtain the hidden state sequence that maximizes the probability of generating the joint observation vector sequence.

[0035] It can be understood that the hidden state sequence output after the joint observation vector sequence is processed by the Viterbi decoding strictly satisfies the topological constraints of the assembly process, that is, the state transitions in the hidden state sequence only occur within the allowed directions defined by the allowed transition direction constraints and the parallel process merging constraints.

[0036] As you can understand, the table below shows the input data format involved in the joint state sequence decoding process of the improved Hidden Markov Model within a sample time window of the automobile generator assembly line, see Table 1.

[0037] Table 1: Correspondence between Joint Observation Vector Sequence and Decoded Output Hidden State Sequence In one embodiment of the present invention, the specific implementation of identifying the actual operating state segment of each assembly station in the automotive generator assembly production line based on the result of joint state sequence decoding processing and comparing the actual operating state segment with a preset standard assembly process segment to generate an assembly deviation mark set is as follows: Based on the hidden state corresponding to each time point in the result of the joint state sequence decoding processing, the hidden state is mapped to the assembly process node name in the automotive generator assembly production line; the hidden state sequence is traversed in chronological order; and time intervals with multiple consecutive time points and unchanged hidden states are extracted as an actual operating state segment. Each actual operating state segment includes a start timestamp, an end timestamp, and a... The corresponding assembly process node name is used to obtain the standard assembly process segment set of the automotive generator assembly production line. The standard assembly process segment set includes the standard start timestamp, standard end timestamp, and standard duration interval for each assembly process node. The actual operating state segment is compared with the standard assembly process segment with the same assembly process node name in the standard assembly process segment set. The difference between the actual duration interval of the actual operating state segment and the standard duration interval is calculated. When the absolute value of the difference exceeds a preset duration deviation threshold, a duration deviation mark corresponding to the assembly process node is generated and the duration deviation mark is added to the assembly deviation mark set.

[0038] The specific implementation of constructing a deviation-material association graph for each assembly station based on the assembly deviation mark set and the component material traceability data is as follows: The deviation-material association graph is used to describe the mapping relationship between the assembly deviation type and the batch of components used. Component material traceability data is extracted from the original data. The component material traceability data includes the batch number of components used in each assembly cycle at each assembly station and the supplier information corresponding to each component batch number. Each duration deviation mark in the assembly deviation mark set is associated with the assembly process node that generated the duration deviation mark. Then, the corresponding assembly station and the corresponding assembly cycle are located through the assembly process node. The batch number of components used by the assembly station in the assembly cycle is obtained. The component batch number is used as the material node in the deviation-material association graph, and the duration deviation mark is used as the deviation node in the deviation-material association graph. An association edge is established between the deviation node and the material node. The weight of the association edge is equal to the frequency of the deviation node and the material node appearing together in the same assembly cycle. All deviation nodes, all material nodes, and all association edges on the same assembly station are aggregated to generate the deviation-material association graph corresponding to the assembly station.

[0039] In specific implementation, taking an assembly cycle of an automotive generator assembly production line as an example, the process of identifying actual operating state segments based on the results of joint state sequence decoding and comparing them with standard assembly process segments to generate an assembly deviation mark set is as follows: Based on the hidden state corresponding to each time point in the results of joint state sequence decoding, the hidden state is mapped to the name of the assembly process node; the hidden state sequence is traversed in chronological order, and the time interval with multiple consecutive time points and unchanged hidden states is extracted as an actual operating state segment. Each actual operating state segment contains a start timestamp, an end timestamp, and the corresponding assembly process node name.

[0040] In some embodiments, a set of standard assembly process segments is obtained, which includes a standard start timestamp, a standard end timestamp, and a standard duration interval for each assembly process node. The actual running state segments are compared with the standard assembly process segments in the set of standard assembly process segments that have the same assembly process node name, and the difference between the actual duration interval and the standard duration interval is calculated. The formula for calculating the difference is: in: This represents the difference between the actual duration and the standard duration corresponding to assembly process node p. This represents the length of the actual duration interval of the actual running state segment corresponding to assembly process node p. This represents the standard duration interval length corresponding to assembly process node p. Optionally, when the absolute value of the difference exceeds a preset duration deviation threshold, a duration deviation marker corresponding to the assembly process node is generated, and this duration deviation marker is added to the assembly deviation marker set.

[0041] In specific implementation, the process of constructing a deviation-material association map for each assembly station based on the assembly deviation mark set and component material traceability data is as follows: The deviation-material association map is used to describe the mapping relationship between the assembly deviation type and the batch of components used; Component material traceability data is extracted from the original data, which includes the batch number of components used in each assembly cycle at each assembly station and the supplier information corresponding to each batch number; Each duration deviation mark in the assembly deviation mark set is associated with the assembly process node that generated the mark, and then the corresponding assembly station and the corresponding assembly cycle are located through the assembly process node.

[0042] In some embodiments, the batch number of the parts used by the assembly station during the assembly cycle is obtained, and the batch number of the parts is used as a material node in the deviation-material association graph, and the duration deviation marker is used as a deviation node in the deviation-material association graph; an association edge is established between the deviation node and the material node, and the weight of the association edge is equal to the frequency of the deviation node and the material node appearing together in the same assembly cycle; all deviation nodes, all material nodes and all association edges on the same assembly station are aggregated to generate the deviation-material association graph corresponding to the assembly station.

[0043] It can be understood that the deviation-material association graph is a weighted bipartite graph, where the deviation node set and the material node set are connected by association edges, and each association edge carries an integer weight. Referring to Table 2, the data on nodes and association edges in the deviation-material association graph generated by an assembly station within one assembly cycle are provided.

[0044] Table 2: Deviation-Material Association Graph Nodes and Association Edges Data Table In one embodiment of the present invention, the specific implementation of performing frequent subgraph mining processing on the deviation-material association graph to extract the abnormal association patterns that recur across assembly stations is as follows: All deviation-material association graphs corresponding to all assembly stations are merged into a total deviation-material association graph. The total deviation-material association graph includes all deviation nodes, all material nodes, and all associated edges. A subgraph support threshold is set, and all subgraph structures in the total deviation-material association graph whose number of nodes is not greater than a preset maximum subgraph size are traversed. The frequency of each subgraph structure appearing in the total deviation-material association graph is calculated. Subgraph structures whose frequency is greater than or equal to the subgraph support threshold are selected as frequent subgraph candidate sets. For each frequent subgraph in the frequent subgraph candidate set, the deviation type corresponding to the deviation node and the part batch number corresponding to the material node are extracted. The correspondence between the deviation type and the part batch number is taken as an abnormal association pattern. The abnormal association patterns are sorted from high to low according to the frequency of their corresponding frequent subgraphs appearing in the total deviation-material association graph. The abnormal association patterns with the highest ranking are extracted as abnormal association patterns that recur across assembly stations.

[0045] The system also includes a root cause tracing module, which is used to perform root cause tracing processing on the abnormal association patterns and generate a root cause tracing report.

[0046] The specific implementation involves obtaining the target component batch number from the abnormal association pattern that recurs across assembly stations, querying the target supplier information and target production date information corresponding to the target component batch number from the component material traceability data, extracting all assembly deviation markers associated with the target component batch number from the original data, grouping all assembly deviation markers according to assembly station and assembly time to obtain a deviation distribution matrix, performing principal component analysis dimensionality reduction on the deviation distribution matrix to extract the principal component direction with the highest contribution rate, marking the assembly station whose load coefficient on the principal component direction exceeds a preset load threshold as a high-risk station, constructing a root cause tracing path according to the target supplier information, target production date information, and location information of the high-risk station in chronological and station order, the root cause tracing path starts from the target component batch number, passes through the high-risk station, and reaches the corresponding assembly deviation marker, and outputting the root cause tracing path as a root cause tracing report to the digital management interface of the automotive generator assembly production line.

[0047] In specific implementation, taking an automotive generator assembly line as an example, the process of extracting recurring abnormal association patterns across assembly stations by performing frequent subgraph mining on the deviation-material association graph is as follows: All deviation-material association graphs corresponding to all assembly stations are merged into a single overall deviation-material association graph, which includes all deviation nodes, all material nodes, and all associated edges; a subgraph support threshold is set, and all subgraph structures in the overall deviation-material association graph with a node count not exceeding a preset maximum subgraph size are traversed, calculating the support of each subgraph structure in the overall deviation-material association graph. The frequency of occurrence is used to select subgraph structures whose frequency is greater than or equal to the subgraph support threshold as a frequent subgraph candidate set. For each frequent subgraph in the frequent subgraph candidate set, the deviation type corresponding to the deviation node and the component batch number corresponding to the material node are extracted. The correspondence between the deviation type and the component batch number is used as an abnormal association pattern. The abnormal association patterns are sorted from high to low according to the frequency of their corresponding frequent subgraphs in the overall deviation-material association graph. The abnormal association patterns with the highest ranking are extracted as abnormal association patterns that recur across assembly stations.

[0048] In some embodiments, the system further includes a root cause tracing module, used to perform root cause tracing processing on the abnormal association patterns and generate a root cause tracing report; obtain the target component batch number in the abnormal association patterns that recur across assembly stations, and query the target supplier information and target production date information corresponding to the target component batch number from the component material traceability data; extract all assembly deviation markers associated with the target component batch number from the raw data, group all assembly deviation markers according to assembly station and assembly time to obtain a deviation distribution matrix; perform principal component analysis dimensionality reduction processing on the deviation distribution matrix, extract the principal component direction with the highest contribution rate, and mark the assembly station whose load coefficient on the principal component direction exceeds a preset load threshold as a high-risk station; construct a root cause tracing path according to the target supplier information, the target production date information, and the location information of the high-risk station, in chronological order and station order, the root cause tracing path starting from the target component batch number, passing through the high-risk station, and reaching the corresponding assembly deviation marker; output the root cause tracing path as a root cause tracing report to the digital management interface of the automotive generator assembly production line.

[0049] Optionally, the formula for calculating the loading coefficients in the principal component analysis dimensionality reduction process is as follows: in: Let f represent the load coefficient of the f-th assembly station in the w-th principal component direction, where f is the assembly station index ranging from 1 to F, F is the total number of assembly stations, and w is the principal component direction index, with w=1 indicating the principal component direction with the highest contribution rate. This represents the f-th component in the eigenvector of the w-th principal component direction. For summation index variables, This represents the sum of squares of all components of the eigenvector in the direction of the w-th principal component.

[0050] It is understood that the root cause tracing path is a directed path, with the starting node being the target component batch number, the intermediate node being the high-risk workstation, and the ending node being the assembly deviation mark.

[0051] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A digital management system for automotive generator assembly production lines based on big data, characterized in that: The system includes: The data preprocessing module acquires raw data including sensor sequence data, assembly equipment status parameter data, and component material traceability data. It performs time alignment and anomaly marking on the sensor sequence data to generate a regularized sensor data stream. It also performs status coding and change point detection on the assembly equipment status parameter data to generate a regularized equipment status stream. The state decoding analysis module uses an improved hidden Markov model to perform joint state sequence decoding processing on the regularized sensor data stream and the regularized equipment state stream. The improved hidden Markov model constructs a state transition matrix based on the topological constraints of the assembly process. The assembly deviation analysis module identifies the actual operating state segments of each assembly station in the automotive generator assembly production line based on the results of the joint state sequence decoding process, and compares the actual operating state segments with the preset standard assembly process segments to generate an assembly deviation mark set. The correlation graph construction module constructs a deviation-material correlation graph for each assembly station based on the assembly deviation mark set and the component material traceability data. The anomaly pattern mining module performs frequent subgraph mining processing on the deviation-material association graph, extracts the abnormal association patterns that recur across assembly stations, and pushes them to the digital management interface of the automotive generator assembly production line.

2. The big data-based automobile generator assembly production line digital management system according to claim 1, characterized in that, The improved Hidden Markov Model constructs a state transition matrix based on the topological constraints of the assembly process, including: Obtain the standard assembly process sequence of the automotive generator assembly production line. The standard assembly process sequence includes multiple assembly process nodes arranged in the order of assembly and the process dependencies between each assembly process node. Each assembly process node is mapped to a hidden state of the improved Hidden Markov Model, and the process dependencies are mapped to allowable transition direction constraints between hidden states, wherein the previous process node can only transfer to the next process node or remain at the current process node, and transfer from the next process node to the previous process node is prohibited. An initial state transition matrix is ​​constructed based on the allowed transition direction constraints. The matrix elements corresponding to the allowed transition directions in the initial state transition matrix are set as variable weight parameters to be trained, and the matrix elements corresponding to the prohibited transition directions in the initial state transition matrix are fixed to zero values. The parallel process merging constraint contained in the assembly process topology constraint is incorporated into the initial state transition matrix. The parallel process merging constraint is used to indicate that multiple parallel process nodes are allowed to transition to each other during execution and have the same transition probability. The initial state transition matrix after being processed by the allowed transition direction constraint and the parallel process merging constraint is subjected to row normalization to generate a state transition matrix that satisfies the assembly process topology constraint.

3. The digital management system for automotive generator assembly line based on big data as described in claim 1, characterized in that, Perform time alignment and anomaly marking on the sensor sequence data to generate a regularized sensor data stream, including: Sensor sequence data for each data acquisition node is extracted from the raw data. The sensor sequence data includes torque sensor numerical sequences, displacement sensor numerical sequences, and vibration sensor numerical sequences. The timestamp fields corresponding to the torque sensor numerical sequence, the displacement sensor numerical sequence, and the vibration sensor numerical sequence are identified. The three sequences are resampled according to the same time granularity using the timestamp fields as the alignment reference to obtain a time-aligned multi-channel sensor matrix. Perform sliding window statistical processing on each channel in the time-aligned multi-channel sensor matrix to calculate the local mean and local standard deviation of the sensor values ​​within each sliding window; Each sensor value within each sliding window is compared with the local mean and the local standard deviation. When a sensor value exceeds the range of the local mean plus or minus three times the local standard deviation, the sensor value is marked as an outlier, and the outlier value is replaced with the local median value to generate a regularized sensor data stream.

4. The big data-based automobile generator assembly production line digital management system according to claim 1, characterized in that, The process of performing state coding and change point detection on the assembly equipment state parameter data to generate a regularized equipment state stream includes: The assembly equipment status parameter data for each assembly equipment is extracted from the raw data. The assembly equipment status parameter data includes a sequence of equipment operating current parameters, a sequence of equipment operating temperature parameters, and a sequence of equipment operating speed parameters. Each current value in the equipment operating current parameter sequence is mapped to a discrete current state code according to a preset current state partition table; each temperature value in the equipment operating temperature parameter sequence is mapped to a discrete temperature state code according to a preset temperature state partition table; and each speed value in the equipment operating speed parameter sequence is mapped to a discrete speed state code according to a preset speed state partition table. The discrete current state code, the discrete temperature state code, and the discrete speed state code at the same time point are concatenated to generate a combined state code sequence corresponding to the time point. The combined state code sequence is subjected to cumulative sum and change point detection processing. The cumulative sum statistic of the combined state code at each time point is calculated. When the cumulative sum statistic of multiple consecutive time points exceeds the preset change threshold, the time point is marked as a device state change point. The combined state code sequence between two adjacent device state change points is taken as a steady-state operation segment. The combined state code that appears most frequently in each steady-state operation segment is taken as the representative state code of the steady-state operation segment. The representative state codes of all steady-state operation segments are arranged in chronological order to generate a regular device state flow.

5. The big data based automobile generator assembly production line digital management system according to claim 2, characterized in that, The process of using an improved Hidden Markov Model to perform joint state sequence decoding of the regularized sensor data stream and the regularized device state stream includes: The regularized sensor data stream is used as the observation sequence input of the improved hidden Markov model, and the regularized device state stream is used as the auxiliary observation sequence input of the improved hidden Markov model. Construct the emission probability matrix of the improved hidden Markov model, where each element in the emission probability matrix represents the joint generation probability of the observation sequence and the auxiliary observation sequence in a given hidden state; Each observation vector in the observation sequence is concatenated with each auxiliary observation vector in the auxiliary observation sequence to generate a joint observation vector sequence; Based on the emission probability matrix and the state transition matrix that satisfies the topological constraints of the assembly process, Viterbi decoding is performed on the joint observation vector sequence to solve for the hidden state sequence that maximizes the generation probability of the joint observation vector sequence. The hidden state sequence is output as the result of decoding the joint state sequence, and each hidden state in the hidden state sequence corresponds to the assembly process node of the automobile generator assembly line at the corresponding time point.

6. The big data-based automobile generator assembly production line digital management system according to claim 5, characterized in that, Based on the results of joint state sequence decoding, the actual operating state segment of each assembly station in the automotive generator assembly production line is identified. This actual operating state segment is then compared with a preset standard assembly process segment to generate an assembly deviation marker set, including: Based on the hidden state corresponding to each time point in the result of the joint state sequence decoding process, the hidden state is mapped to the assembly process node name in the automobile generator assembly production line. Traverse the hidden state sequence in chronological order, extract the time intervals where multiple consecutive time points and hidden states remain unchanged as an actual running state segment, and each actual running state segment contains a start timestamp, an end timestamp and the corresponding assembly process node name. Obtain a set of standard assembly process segments for the automotive generator assembly production line. The set of standard assembly process segments includes a standard start timestamp, a standard end timestamp, and a standard duration range for each assembly process node. The actual running state segment is compared with the standard assembly process segment with the same assembly process node name in the standard assembly process segment set, and the difference between the actual duration interval of the actual running state segment and the standard duration interval is calculated. When the absolute value of the difference exceeds the preset duration deviation threshold, a duration deviation mark corresponding to the assembly process node is generated, and the duration deviation mark is added to the assembly deviation mark set.

7. The big data based automobile generator assembly production line digital management system according to claim 1, characterized in that, Based on the assembly deviation marker set and the component material traceability data, a deviation-material correlation graph is constructed for each assembly station, including: The deviation-material correlation diagram is used to describe the mapping relationship between assembly deviation types and batches of parts used; Extract component material traceability data from the raw data. The component material traceability data includes the component batch number used in each assembly cycle at each assembly station and the supplier information corresponding to each component batch number. Each duration deviation mark in the assembly deviation mark set is associated with the assembly process node that generates the duration deviation mark, and then the corresponding assembly station and the corresponding assembly cycle are located through the assembly process node. Obtain the batch number of the parts used by the assembly station during the assembly cycle, use the batch number of the parts as a material node in the deviation-material association diagram, and use the duration deviation mark as a deviation node in the deviation-material association diagram; Establish an association edge between the deviation node and the material node, and the weight of the association edge is equal to the frequency of the deviation node and the material node appearing together in the same assembly cycle; Aggregate all deviation nodes, all material nodes, and all associated edges on the same assembly station to generate a deviation-material association graph corresponding to the assembly station. 8.The big data-based automobile generator assembly production line digital management system according to claim 7, characterized in that, Frequent subgraph mining is performed on the deviation-material association graph to extract abnormal association patterns that recur across assembly stations, including: All deviation-material association maps corresponding to assembly stations are merged into a single overall deviation-material association map, which includes all deviation nodes, all material nodes, and all associated edges. Set a subgraph support threshold, traverse all subgraph structures in the overall deviation-material association graph whose number of nodes is not greater than the preset maximum subgraph size, and calculate the frequency of each subgraph structure in the overall deviation-material association graph. Subgraph structures whose frequency is greater than or equal to the subgraph support threshold are selected as frequent subgraph candidate sets. For each frequent subgraph in the frequent subgraph candidate set, extract the deviation type corresponding to the deviation node and the component batch number corresponding to the material node contained in the frequent subgraph, and use the correspondence between the deviation type and the component batch number as the abnormal association mode. The abnormal association patterns are sorted from high to low according to the frequency of their corresponding frequent subgraphs in the overall deviation-material association diagram, and the abnormal association patterns with the highest ranking are extracted as abnormal association patterns that recur across assembly stations. 9.The big data based automobile generator assembly production line digital management system according to claim 1, wherein, The system also includes The root cause tracing module is used to perform root cause tracing processing on the abnormal association patterns and generate a root cause tracing report, specifically including: Obtain the target component batch number in the abnormal association pattern that occurs repeatedly across assembly stations, and query the target supplier information and target production date information corresponding to the target component batch number from the component material traceability data; Extract all assembly deviation markers associated with the batch number of the target component from the original data, and group all assembly deviation markers according to assembly station and assembly time to obtain a deviation distribution matrix; Principal component analysis is performed on the deviation distribution matrix to reduce its dimensionality, and the principal component directions with the highest contribution rates are extracted. Assembly stations whose load coefficients on the principal component directions exceed a preset load threshold are marked as high-risk stations. Based on the target supplier information, the target production date information, and the location information of the high-risk workstation, a root cause tracing path is constructed according to the time sequence and workstation sequence. The root cause tracing path starts from the target component batch number, passes through the high-risk workstation, and arrives at the corresponding assembly deviation mark. The root cause tracing path is output as a root cause tracing report to the digital management interface of the automotive generator assembly line.

10. The big data based automobile generator assembly production line digital management system according to claim 5, characterized in that, Based on the emission probability matrix and the state transition matrix satisfying the assembly process topology constraints, Viterbi decoding is performed on the joint observation vector sequence to solve for the hidden state sequence that maximizes the generation probability of the joint observation vector sequence, including: Initialize the Viterbi path probability matrix, input the first joint observation vector of the joint observation vector sequence into the emission probability matrix, calculate the probability value of each hidden state at the initial time, and set the predecessor state corresponding to each hidden state to empty; Starting from the second joint observation vector in the joint observation vector sequence, perform iterative calculation processing on each joint observation vector in the joint observation vector sequence. For each hidden state, calculate the product of the transition probability from all hidden states in the previous time step to the current hidden state and the corresponding path probability in the previous time step. The maximum value is selected from the product as the cumulative path probability of the current hidden state at the current time, and the hidden state from the previous time that generated the maximum value is recorded as the predecessor state of the current hidden state. The cumulative path probability and the predecessor state are recorded in the corresponding positions of the Viterbi path probability matrix. After the iterative calculation is completed, the hidden state with the largest cumulative path probability at the last time step in the Viterbi path probability matrix is ​​selected as the starting point of the final state. Starting from the initial point of the final state, based on the preceding state information recorded in the Viterbi path probability matrix, we backtrack to the initial time to obtain the hidden state sequence that maximizes the probability of generating the joint observation vector sequence.

Images