Industrial big data analysis method for micro inverter manufacturing

Abstract

The invention relates to the technical field of industrial big data processing, and discloses an industrial big data analysis method for micro inverter manufacturing, and the method comprises the steps: analyzing a discrete event log of a manufacturing execution system to construct a directed traceability topological graph; defining a relative process time window and intercepting continuous time sequence data based on the process entry and exit events; calculating a first-order derivative sequence and identifying a morphological segmentation anchor point, dividing continuous time series data into functional sub-phases and calculating local statistical characteristics; splicing each phase feature to generate a structured process fingerprint vector, and writing the structured process fingerprint vector into a topological graph connection edge; based on failure nodes, frequent isomorphic subgraphs are reversely mined to lock a failure mode, and logic alignment of heterogeneous data and dimensionality reduction projection of a waveform form are realized on the premise of not depending on clock synchronization through a relative event boundary and derivative form segmentation mechanism; the technical problem that local process anomaly is covered by cross-system data space-time dislocation and statistical characteristics is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to an industrial big data analysis method for micro-inverter manufacturing, and belongs to the technical field of industrial big data processing. BACKGROUND

[0002] The current micro-inverter is a precision power electronic device, and the manufacturing process involves multiple discrete processes such as surface mounting, glue filling and aging test, accompanied by continuous physical quantity changes such as temperature curve and pressure waveform. The existing manufacturing data management system has two data modes: the manufacturing execution system generates structured logs based on discrete events, and the industrial internet of things generates unstructured continuous time series data based on absolute clocks. For such heterogeneous data, conventional processing relies on absolute timestamp alignment, assuming that different subsystem time references are strictly synchronized, and correlating discrete work order records by querying sensor values at a specific time point. This method is suitable for single closed and highly unified clock automated production lines, and meets the basic production monitoring and traceability requirements.

[0003] In large-scale discrete manufacturing scenarios, the manufacturing execution system and the industrial internet of things system operate on different network levels and hardware architectures, and the local clock of each device controller and the clock of the server system have difficult-to-eliminate drift and delay. When micro-inverter failure is caused by cross-process, multi-factor nonlinear coupling, simple absolute time alignment faces failure risks, and the existing technology has the following shortcomings: 1. Lack of a unified microsecond-level synchronization clock source, resulting in phase misalignment based on absolute timestamp data correlation, causing sensor-captured abnormal waveform fragments to be incorrectly attributed to adjacent normal work orders, resulting in noisy training sample labels; 2. Traditional global statistical feature extraction methods ignore the morphological semantics of signals in different process stages, and cannot identify hidden defects with normal mean but locally chaotic time sequence logic; 3. The existing static traceability graph model only records material level relationships, cutting off the time sequence correlation of continuous production process state evolution of the same device, making it difficult to discover complex failure modes induced by device progressive degradation or specific processing sequence combinations.

[0004] Therefore, how to construct a data analysis method that aligns heterogeneous discrete events and continuous signals independently of the absolute system clock and extracts cross-process time evolution feature data has become a technical problem to be solved by the present application. SUMMARY

[0005] To solve the problems raised in the background art, the technical solution of the present application is as follows: an industrial big data analysis method for micro-inverter manufacturing, comprising the following steps: Analyzing the discrete event logs of the manufacturing execution system, taking the single product unique identification code as the core node, and constructing a directed traceability topology graph containing directed connection edges pointing to the device node and the material batch node; In response to the process entry events and process exit events recorded in the discrete event logs, the relative process time window of the single product in a specific process is defined; obtaining equipment continuous time series data corresponding to the process, intercepting the equipment continuous time series data within the relative process time window, calculating a first derivative sequence of the equipment continuous time series data with respect to time; identifying a time point at which the sign of the value in the first derivative sequence is reversed or the value crosses a preset change rate threshold, and defining the time point as a morphology segmentation anchor point; logically dividing the equipment continuous time series data into a plurality of continuous functional sub-phases according to the morphology segmentation anchor points, and respectively calculating local statistical features of each functional sub-phase; splicing the local statistical features of all functional sub-phases according to the time sequence logic to generate a structured process fingerprint vector, and writing the structured process fingerprint vector as an attribute value into a directed connection edge between a corresponding equipment node and a single product node in the directed traceability topology graph, to generate a heterogeneous manufacturing process graph; obtaining single product final test failure data, performing reverse path traversal in the heterogeneous manufacturing process graph with the failure single product node as the terminal point, extracting a candidate path containing equipment nodes, material batch nodes and structured process fingerprint vectors, and mining a subgraph structure of common characteristic attributes appearing with a frequency exceeding a preset support threshold in the candidate path, to determine the subgraph structure as a failure cause mode.

[0006] Preferably, the step of logically dividing the equipment continuous time series data into a plurality of continuous functional sub-phases according to the morphology segmentation anchor points comprises: locating a time point at which the first value is greater than a positive change rate threshold in the first derivative sequence, defining the time point as a rising edge start anchor point; locating a time point at which the first value absolute value is less than a steady state change rate threshold and the duration exceeds a preset judgment time, defining the time point as a platform steady state anchor point; locating a time point at which the first value is less than a negative change rate threshold, defining the time point as a falling edge start anchor point; and based on the time sequence positions of the rising edge start anchor point, the platform steady state anchor point and the falling edge start anchor point, dividing the equipment continuous time series data into a temperature rising and pressure increasing phase, a constant temperature and pressure maintaining phase and a temperature decreasing and pressure releasing phase.

[0007] Preferably, the step of respectively calculating the local statistical features of each functional sub-phase comprises: for each functional sub-phase, calculating the arithmetic mean, the standard deviation and the peak value of the data points in the phase as basic statistical quantities; calculating the duration proportion of the functional sub-phase within the relative process time window as a time domain feature quantity; and combining the basic statistical quantities and the time domain feature quantity to generate a local feature sub-vector of the functional sub-phase.

[0008] Preferably, the step of concatenating the local statistical features of all functional sub-phases according to the time sequence to generate the structured process fingerprint vector comprises: connecting the local feature sub-vectors corresponding to each functional sub-phase in series according to the order of occurrence of the functional sub-phases in the physical process; and performing normalization processing on the connected vector to generate a structured process fingerprint vector with fixed dimensions, so that the vector represents the topological evolution form of the device continuous time sequence data in the time dimension.

[0009] Preferably, the step of constructing the directed traceability topological graph further comprises constructing a time sequence adjacency chain for the same device node, specifically comprising: performing time sequence sorting on all single-product nodes processed by the same device node according to the process end time recorded in the discrete event log; establishing a directed time sequence adjacency edge between the adjacent pre-sequence single-product node and the post-sequence single-product node; calculating the difference feature vector ΔV between the structured process fingerprint vector of the post-sequence single-product node and the structured process fingerprint vector of the pre-sequence single-product node, and the calculation rule of the difference feature vector ΔV is: ΔV = V n -V n-1 , wherein V n is the structured process fingerprint vector of the post-sequence single-product node, and V n-1 is the structured process fingerprint vector of the pre-sequence single-product node; and mapping the difference feature vector ΔV as an evolution attribute representing the dynamic drift of the device state to the directed time sequence adjacency edge.

[0010] Preferably, the step of mining sub-graph structures with high-frequency common feature attributes in the candidate paths comprises: inputting all extracted candidate paths into a frequent sub-graph mining algorithm; counting the occurrence frequency of each sub-graph structure in the candidate path set; identifying common sub-graph structures with an occurrence frequency exceeding a preset support threshold; and when a specific structured process fingerprint vector attribute interval is included in the common sub-graph structure, combining the specific attribute interval with the corresponding device node to lock as the root cause mode leading to failure.

[0011] Preferably, the step of defining the relative process time window of the single product in the specific process further comprises performing logical alignment of multi-source data: when the sampling frequency of the device continuous time sequence data is higher than the time resolution of the discrete event log, taking the time of the process entry event as the relative zero point; uniformly mapping the time axis of the device continuous time sequence data to the offset time axis relative to the relative zero point; and eliminating the data segments in the offset time axis after the process exit event occurs, and retaining the effective data segments within the closed interval of the process entry event and the process exit event.

[0012] Preferably, the step of writing the structured process fingerprint vector as an attribute value into the directed connection edge between the corresponding equipment node and the single product node in the directed provenance topology graph further comprises: obtaining the unique identification code of the auxiliary fixture used in the process and the identity code of the current operator as discrete attributes, and writing them into the directed connection edge between the equipment node and the single product node; and making the directed connection edge simultaneously carry the continuous quantity feature representing the physical process form and the discrete quantity feature representing the production resource configuration.

[0013] Preferably, the method further comprises the step of performing active blocking based on the mined failure cause mode: real-time monitoring of the generated heterogeneous manufacturing process graph in the subsequent production process; calculating the similarity of the connection attribute or subgraph structure between the newly generated single product node and the equipment node and the failure cause mode; and when the similarity exceeds the preset safety threshold, outputting the shutdown inspection instruction or parameter correction instruction for the equipment.

[0014] Preferably, in the step of analyzing the discrete event log of the manufacturing execution system, the discrete event log contains the single product serial number, the process name, the process start timestamp, the process end timestamp, the processing equipment number, and the material batch number.

[0015] Compared with the prior art, the present application has the following advantages: 1. The process start and end anchor points are determined by analyzing the discrete event log of the manufacturing execution system, the anchor points are used as the logical boundaries of the intercepted equipment continuous time sequence data, the relative process time window-based heterogeneous data projection mechanism is constructed, the logical sequence of process events is used to replace the absolute clock physical synchronization of the equipment, and in the case that there is clock drift or inconsistent sampling frequency between the manufacturing subsystem and the industrial internet of things system, the deterministic mapping relationship between the discrete work order node and the continuous physical signal feature in the graph database is established, the data correlation failure caused by the deviation of the cross-system time and space reference is eliminated, and the logical topology level consistency and integrity of the heterogeneous manufacturing data are ensured.

[0016] 2. The first derivative sequence of the intercepted time sequence data is calculated, and the sign flip or threshold crossing point is identified as a form segmentation anchor point, the single-dimensional continuous waveform data is adaptively divided into multiple sub-phases with independent physical semantic functions, the structured process fingerprint vector is generated by sequentially splicing the phases, and the feature extraction process is automatically adapted to the process period fluctuation caused by equipment component wear or control parameter fine tuning based on the adaptive segmentation logic of the internal form change of the signal, without pre-setting a fixed time template to avoid phase misplacement interference, the local evolution characteristics of the waveform data in different process stages are retained, and the global statistical indicators that cannot represent the stage process abnormalities are revealed.

[0017] 3. Introducing process end time based time-adjacent edges in the provenance topology graph, concatenating single-product nodes processed by the same equipment in production sequence, calculating process fingerprint vector difference features between adjacent nodes along time-adjacent edges, and reconstructing the topology structure to embed dynamic equipment state evolution dimension in static material provenance relationship, so that the graph mining algorithm analyzes the cumulative trend or mutation in the difference feature time series, identifies the heat accumulation caused by continuous production, the gradual accumulation of auxiliary material residues, or the hidden failure mode induced by specific processing sequence combination, and makes up for the dimensional deficiency of traditional static graph model in representing process time variation. BRIEF DESCRIPTION OF DRAWINGS

[0018] Fig. 1 The micro-inverter big data analysis flowchart based on the relative process time window and the morphology segmentation of the present application; Fig. 2 The sensitivity analysis and optimal threshold selection diagram of the similarity safety threshold and failure identification performance index of the present application; Fig. 3 The system architecture and data flow principle diagram of the present application integrating data logic alignment and graph mining services. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0020] The application provides an industrial big data analysis method for manufacturing micro-inverters, which analyzes a manufacturing execution system discrete event log, takes a single product unique identification code as a core node, constructs a directed traceability topology graph containing directed connection edges pointing to equipment nodes and material batch nodes, and the discrete event log contains a single product serial number, a process name, a process start timestamp, a process end timestamp, a processing equipment number, and a material batch number. For the same equipment node, according to the discrete event log record process end time, all single product nodes processed by the equipment node are executed in time sequence sorting, and a directed time sequence adjacency edge is established between adjacent previous single product nodes and subsequent single product nodes in the sorting. In response to the process entry event and the process exit event recorded in the discrete event log, the relative process time window of the single product in the specific process is defined. For the case where the sampling frequency of the equipment continuous time series data is higher than the time resolution of the discrete event log, the moment when the process entry event occurs is taken as the relative zero point t0, the equipment continuous time series data time axis is uniformly mapped to the offset time axis t' relative to the relative zero point t0, the data segment located after the moment when the process exit event occurs in the offset time axis t' is removed, the effective data segment within the closed interval of the process entry event and the process exit event is retained, and the multi-source data logic alignment is realized. The corresponding equipment continuous time series data of the process is obtained, the equipment continuous time series data within the relative process time window is intercepted, the first order derivative sequence S' of the equipment continuous time series data is calculated, the moment when the numerical symbol of the first order derivative sequence S' changes or the numerical value crosses the preset change rate threshold is identified, the moment is defined as a morphological segmentation anchor point, and the equipment continuous time series data is logically divided into multiple continuous functional sub-phases according to the morphological segmentation anchor point. For example, in the glue filling and curing furnace temperature control process, the first moment when the numerical value is greater than the positive change rate threshold such as 2.0 DEG C / s is located in the first order derivative sequence of the temperature time series data, and is defined as the rising edge start anchor point. The first moment when the numerical absolute value is less than the steady state change rate threshold such as 0.5 DEG C / s and the duration exceeds the preset determination time such as 30s is located, and is defined as the platform steady state anchor point. The first moment when the numerical value is less than the negative change rate threshold such as -1.5 DEG C / s is located, and is defined as the falling edge start anchor point. Based on the time sequence positions of the rising edge start anchor point, the platform steady state anchor point and the falling edge start anchor point, the equipment continuous time series data is divided into a temperature rising and pressure increasing phase, a constant temperature and pressure maintaining phase, and a temperature decreasing and pressure releasing phase.

[0021] respectively, and the arithmetic mean, standard deviation, and peak value of the data points in each functional sub-phase are calculated as the basic statistical quantities; the duration ratio of the functional sub-phase in the relative process time window is calculated as the time domain feature quantity; the basic statistical quantities and the time domain feature quantity are combined to generate a local feature sub-vector of the functional sub-phase; the local feature sub-vectors corresponding to each functional sub-phase are serially connected in the order of occurrence of the functional sub-phase in the physical process, and normalization processing is performed on the connected vector to generate a structured process fingerprint vector V with fixed dimension; the structured process fingerprint vector V is written as an attribute value into the directed connection edge between the corresponding device node and the single product node in the directed traceability topology graph, and a heterogeneous manufacturing process graph is generated; the structured process fingerprint vector V of the subsequent single product node is calculated n and the structured process fingerprint vector V of the previous single product node n-1 , and the difference feature vector AV is calculated as AV = V n -V n-1 , the difference feature vector AV is mapped as a dynamic drift evolution attribute representing the state of the device to the directed time sequence adjacency edge, the unique identification code of the auxiliary fixture and the current operator identification code used in the process are obtained, and the unique identification code of the auxiliary fixture and the operator identification code are written as discrete attributes into the directed connection edge between the device node and the single product node, so that the directed connection edge simultaneously carries the continuous feature representing the physical process form and the discrete feature representing the production resource configuration; the single product final test failure data is obtained, and the reverse path traversal is performed in the heterogeneous manufacturing process graph with the failure single product node as the terminal point, the candidate path containing the device node, the material batch node and the structured process fingerprint vector is extracted, and all the extracted candidate paths are input into the frequent subgraph mining algorithm, the frequency of each subgraph structure in the candidate path set is counted, and the common subgraph structure with a frequency exceeding a preset support threshold such as is identified, when a specific structured process fingerprint vector attribute interval is included in the common subgraph structure, the specific attribute interval and the corresponding device node are combined to lock the failure root cause mode, and based on the mined failure cause mode, active blocking is performed, and the generated heterogeneous manufacturing process graph in the subsequent production process is monitored in real time, the similarity between the connection attribute or subgraph structure between the newly generated single product node and the device node and the failure cause mode is calculated, and when the similarity exceeds a preset safety threshold such as 0.85, the shutdown inspection instruction or parameter correction instruction for the device is output.

[0022] Embodiment 1: The micro-inverter manufacturing workshop production line faces the dual challenges of spatial and temporal dislocation of cross-system data and difficulty in locking implicit failure modes. The manufacturing execution system (MES) records work order status in a discrete event-driven mode, and the industrial internet of things (IIoT) platform continuously collects device physical parameters based on an absolute clock. There is a clock drift between the device controller and the server system of the two production lines that cannot be completely eliminated, and the physical process time of wave soldering and glue filling fluctuates with the ambient temperature, resulting in frequent phase dislocation of the traditional alignment method based on absolute time stamp, and the normal process temperature fluctuation is mistakenly attributed to the production process of adjacent failed products, causing the failure of the subsequent failure root cause analysis model due to training sample label noise. To address this challenge, the technical solution of the present application is deployed in the data analysis platform of the production line. The system analyzes the MES discrete event log, constructs a directed traceability topology graph containing single product nodes, device nodes and material batch nodes, establishes a static material traceability skeleton, and in response to process entry events and process exit events, the system defines the relative process time window of each single product at a specific process, intercepts the corresponding device continuous time series data, and maps the absolute time axis to a time axis t' relative to the process entry time t0. The system logically eliminates the influence of cross-system clock drift, realizes accurate alignment of heterogeneous data, calculates the first derivative sequence S' of the intercepted data, identifies the value sign reversal or threshold crossing time point as the morphology segmentation anchor point, adaptively divides the continuous temperature curve into three functional sub-phases of temperature rise and pressure, constant temperature and pressure, and temperature drop and pressure release, respectively. The system calculates the local statistical features of each sub-phase to generate a structured process fingerprint vector V, which is written into the topology graph connection edge as an attribute value, projects the unstructured waveform morphology into a calculable structured feature in the graph, and calculates the difference feature vector ΔV along the time sequence adjacency edge to capture the dynamic drift of the device state in the continuous production process. In the heterogeneous manufacturing process graph, the system starts from the aging test failed single product node and performs reverse frequent subgraph mining. The algorithm identifies a public subgraph structure containing a specific clamp ID, a glue filling machine constant temperature phase temperature standard deviation > 3.5, and a time series difference feature ΔV showing a cumulative upward trend. This structure reveals an implicit failure mode, i.e. the temperature fluctuation gradually increases in the constant temperature phase due to heat accumulation in the specific clamp during continuous use, eventually causing incomplete product glue filling and curing. The system outputs cooling maintenance instructions for the clamp accordingly, and monitors the similarity of the subgraph structure in real time during subsequent production.

[0023] Example 2: A verification platform is built on a real automated production line containing two key processes of surface mount SMT and potting curing. The platform integrates the data interface of MES system and the IIoT sensor network deployed in the key equipment of wave soldering machine and potting machine. The MES system provides discrete work order transfer records, including single product serial number and process in-out station time stamp. The IIoT network continuously collects the wave soldering zone furnace temperature curve and the potting head glue injection pressure waveform at a sampling frequency of 100 Hz. The test actively superimposes 20 dB Gaussian white noise with a signal-to-noise ratio in the IIoT collected original signal to simulate real industrial field interference. Random communication delay is introduced to simulate cross-system clock drift. The test verifies the relative process time window heterogeneous data alignment mechanism and the morphological segmentation feature extraction method. Compared with the traditional absolute time stamp and global statistical feature method, the accuracy of failure mode identification is superior. Three control groups are designed: control group A uses absolute time stamp data alignment. According to the MES recorded process time period, IIoT data is intercepted, and the global mean and standard deviation of the intercepted segment are calculated as features. Control group B uses the relative process time window alignment of the present application, and calculates the global statistical features without performing morphological segmentation and sub-phase feature extraction. The sample group of the present application uses the complete technical solution of the present application, including relative process time window alignment, first-order derivative morphological segmentation, sub-phase local feature extraction, and time series difference feature calculation. Select 500 samples confirmed by aging test to be failed due to potting bubbles and 500 samples with normal function in historical production data as the test set. For each sample, the feature vector is extracted by the above three methods, and input into the same support vector machine SVM classifier for training and prediction. The key intermediate data and the final classification results are shown in Table 1.

[0024] Table 1: Comparison of feature extraction and failure identification results of different methods Control group A has a serious alignment deviation with an average This resulted in a large amount of non-process noise data being mixed into the feature calculation, leading to a recognition accuracy of only 62.4%. Control group B, by aligning relative time windows, reduced the deviation by an average of 20ms, increasing the recognition rate to 78.1%. This indicates that logical alignment removes irrelevant data, and relying on global statistical features cannot distinguish between two drastically different physical states: a slightly lower temperature throughout the process and a sudden temperature drop during the isothermal stage. The sample group of this invention, through morphological segmentation, deconstructs the continuous waveform into three independent physical processes: heating, isothermal, and cooling. It extracts local features from each stage, particularly the isothermal phase standard deviation feature, which accurately captures minute temperature fluctuations during the potting and curing stage, resulting in a recognition accuracy jump to 94.5%. Gradient experiments show that… As the analog clock drift increased from 0s to 5s, the recognition accuracy of the control group A showed a linear decreasing trend, while the recognition accuracy of the present invention sample group remained stable, only showing a slight decrease when the drift exceeded the process interval time. The results confirm that the method of the present invention does not rely on absolute clock synchronization and has robustness against clock drift. By analyzing the timing differential feature ΔV, the present invention sample group issued several warnings of progressive distortion of the pressure waveform caused by the wear of the dispensing head, proving the dynamic evolution value of the differential feature capture equipment. Comparative experiments confirm that the present invention solves the technical problems of cross-system spatiotemporal misalignment and global feature masking of local anomalies based on relative process time window data alignment and morphological segmentation feature extraction mechanism.

[0025] Example 3: This example combines Figs. 1 to 3 This document describes an industrial big data analysis method for the manufacturing of micro-inverters, such as... Fig. 1 As shown, the process begins with high-frequency sampling of continuous time-series data from the equipment and parsing of discrete event logs from the manufacturing execution system. The equipment data includes temperature and pressure waveforms, while the logs contain item serial numbers, process timestamps, and material batches. In response to process entry / exit events, the relative process time window is defined with the process entry time as the relative zero point t0 to eliminate cross-system clock drift and extract valid data segments. Next, the first derivative sequence of the extracted data is calculated, and sign inversion or threshold crossing points are identified to define morphological segmentation anchors. The data is adaptively divided into functional sub-phases such as heating, isothermal, and cooling, and the mean of the local statistical characteristics of each phase is concatenated. The standard deviation and peak value are used to generate a fixed-dimensional structured process fingerprint vector V. At the same time, a directed traceability topology graph with individual products, equipment and material batches as core nodes is constructed. The structured process fingerprint vector V is written into the connecting edges to generate a heterogeneous manufacturing process graph. The differential feature vector ΔV representing equipment drift is calculated. Discrete resource attributes such as auxiliary fixture ID and operator ID are associated. Finally, based on the frequent isomorphic subgraph algorithm, the failure cause pattern is reversed in the heterogeneous graph to lock the root cause pattern containing equipment, parameters and auxiliary materials. In real-time monitoring, the graph structure similarity is calculated to execute the active blocking operation of output shutdown or parameter correction command.

[0026] like Fig. 2As shown in the figure, the horizontal coordinate represents the similarity safety threshold τ, ranging from 0.50 to 0.98, and the vertical coordinate represents the index value, ranging from 0 to 1.0. The figure includes three change curves, namely the solid line representing the precision, the dashed line representing the recall, and the dotted line representing the F1 score. As the similarity safety threshold τ increases from 0.50 to 0.98, the precision shows a monotonous upward trend, while the recall remains high and stable before τ is 0.86, and then sharply decreases. The F1 score curve reaches a peak at τ of 0.88. Fig. 3 As shown in the figure, the top layer of the system is a big data analysis server cluster, which integrates a data logic alignment module responsible for time window definition and offset mapping, a waveform feature extraction engine responsible for derivative calculation, morphological segmentation and fingerprint generation, and a graph mining and blocking service responsible for subgraph matching, failure cause and instruction issuance. The cluster is connected to a data storage server through a gigabit industrial Ethernet. The data storage server includes an industrial process graph database for storing topological graphs and fingerprint vectors, and a manufacturing log and time series database for storing discrete logs and original waveforms. The bottom layer of the system is composed of an industrial internet of things (IIoT) collection gateway and a manufacturing execution system (MES) client. The IIoT collection gateway is responsible for high-frequency sampling and protocol conversion of the temperature control unit of the wave soldering / reflow soldering equipment and the pressure injection unit of the precision glue filling equipment, and uploads high-frequency time series data streams to the server. The MES client inputs scan codes and auxiliary material information through a workstation operation terminal and a scanning gun, forms a process discrete event stream upload, and the analysis server performs real-time operation based on the above data stream. The generated active blocking or parameter correction instructions are fed back to the bottom layer equipment control unit.

[0027] In example 4, the system solves the interface compatibility problem between continuous time series features and discrete graph mining algorithms, and avoids feature drift caused by differences in physical characteristics of equipment. At the initial stage of method deployment, the system performs initialization parameter calibration and model construction procedures. Select 10,000 historical normal single product structured process fingerprint vectors in the past month on the production line as the benchmark data set. Apply the K-means clustering algorithm to divide the data set into K independent statistical clusters. The value of K is determined by the maximum principle of the silhouette coefficient. The system calculates the centroid vector of each cluster, and defines the centroid as the standard process morphology prototype. For each new structured process fingerprint vector V new, the system calculates the Euclidean distance of all standard process morph prototype, maps it to the nearest prototype corresponding to the discrete state identification code, this procedure collapses the continuous high-dimensional feature space into a finite discrete state set, makes the attribute node in the heterogeneous manufacturing process graph have discrete label attributes that can be efficiently traversed and matched by graph algorithms; For the key similarity safety threshold τ setting in the active blocking mechanism, the system performs a historical backtracking sensitivity analysis procedure instead of empirical numerical setting. The system loads the labeled data set containing known failure cases, adjusts the similarity safety threshold τ in the [0.50, 0.99] interval with a step of 0.01. For each τ value, the system simulates the active blocking logic, and counts the false positives and false negatives generated thereby. The system calculates the F1 score at each step. Experimental data shows that as τ increases, the precision rate increases, and the recall rate decreases sharply after exceeding a certain critical point. The system identifies the τ value when the F1 score reaches the global maximum, which is 0.88 in this specific working condition, and locks it as the optimal safety threshold for the current production cycle of the production line.

[0028] According to the calibration procedure, the system constructs a failure cause mode library for the new product. In actual operation, when the connection subgraph structure between the newly generated single product node and the equipment node is mapped after vector quantization, and the similarity calculation result with the stored cause mode structure in the mode library exceeds 0.88, the system determines that the current production state falls into the high-risk interval. The system triggers a hierarchical response mechanism: when the similarity is in the [0.88, 0.92] interval, the system sends a yellow light warning signal to the MES, marking that the batch of products needs to enter the key retest queue; when the similarity exceeds 0.92, the system directly sends a millisecond-level interrupt instruction to the equipment controller, forcing the current process to be suspended until the abnormal state is confirmed and cleared by manual intervention. This embodiment establishes a clear mapping path from continuous signals to discrete states and optimizes the threshold based on statistical optimal solution, avoiding the black box of algorithm landing process parameters.

[0029] Example 5: This embodiment realizes reliable deployment and long-term stable operation in complex and variable industrial sites, performs standardized offline calibration and data filling procedures, collects full-quantity historical data covering all target processes and typical environmental conditions during the initial system construction, establishes a benchmark feature library containing device-process-parameter three-dimensional mapping relationships, for each type of production equipment, records standard time sequence waveforms under different environmental temperatures, load levels and aging degrees through control variable experiments, extracts each functional sub-phase feature parameter using the first derivative shape segmentation algorithm of the invention, constructs an initial device state fingerprint atlas, for each type of key material, combines offline test data and process expert knowledge to calibrate the range of physicochemical property parameters under different process conditions, generate material attribute lookup table, through procedures, convert discrete engineering experience into structured prior knowledge to fill the initial node attributes of heterogeneous manufacturing process graph, provide a definite reference benchmark and logical starting point for subsequent real-time analysis and failure mode mining; before applying the invention to a new production line or introducing a new product model, strictly perform the pre-deployment calibration procedures, the system runs in silent observation mode after connecting to the new environment, only collects data without triggering control instructions, the continuous time is not less than one complete production cycle, during this period, the system automatically compares the real-time collected device time sequence data with the standard waveforms in the benchmark feature library, calculates the deviation distribution statistics, if the deviation exceeds the pre-set adaptive tolerance, the system starts the automatic calibration algorithm, adjusts the relative process time window boundary and shape segmentation threshold, so that the model adapts to the specific physical characteristics and clock drift mode of the current production line, the system displays the potential failure mode subgraph preliminarily mined through the man-machine interface to the process engineers, and after artificial confirmation or correction, it is included in the formal failure cause mode library, the procedure avoids the risk of model mismatch caused by environmental differences through the observation-calibration-confirmation closed-loop process.

[0030] Example 6: For the new process deployment scene of micro-inverter shell airtightness detection, the system adapts the general shape segmentation algorithm to the pressure test waveform with nonlinear attenuation characteristics, solves the engineering problem that general parameters cannot accurately define the physical boundaries of inflation, pressure retention and exhaust, to determine the process-specific shape segmentation anchor point parameters, perform the standardized parameter initialization calibration procedure, the system collects 200 qualified product pressure time sequence data for this process under standard atmospheric pressure and specified temperature range, constructs a gold sample waveform set W golden , the system calculates the first-order derivative sequence of all waveforms in the set with respect to time , according to the prior knowledge of process action logic time sequence, the derivative sub-sequence fragment corresponding to the steady-state pressure retention stage is intercepted, the system calculates the arithmetic mean μ and standard deviation σ of the absolute values of all data points in the sub-sequence fragment, according to the statistical 3σ rule, the steady-state change rate threshold is set as τ steady =μ+3σ, ensuring that the threshold setting covers 99.7% of the background noise fluctuation range, avoiding false segmentation caused by high-frequency noise of the sensor.

[0031] To determine the thresholds for rising and falling edges, the system analyzes... The peak distribution of derivatives corresponds to the rapid inflation and forced deflation stages. The system calculates the lower quartile Q1 of the positive peak value of the derivative sequence during the inflation stage. pos The threshold for the positive rate of change is set to τ. pos =0.5×Q1 pos ; Calculate the upper quartile Q3 of the negative peak of the derivative sequence during the exhaust phase. neg The threshold for the negative rate of change is set to τ. neg =0.5×Q3 neg The logic is set to ensure the capture of the effective action start point, with a 50% signal attenuation margin to accommodate normal fluctuations in gas source pressure, and a preset judgment duration T is set. th The system is configured to calculate the minimum duration T of all gold sample waveforms between the completion of inflation and the start of deflation. min , will T th Set to 0.8×T min A time filter for filtering transient pressure shocks is constructed. Through a statistical distribution feature calibration process, the system concretizes the abstract morphological segmentation logic into a set of deterministic parameters adapted to specific physical processes, enabling standardized migration and implementation of the algorithm in different physical quantity monitoring scenarios. For the continuous feature vector to discrete graph algorithm adaptation, the system initialization phase executes a vector quantization state space calibration procedure, selecting 10,000 structured process fingerprint vectors of single products that have passed quality verification within historical production cycles as the benchmark dataset. The dataset is divided using the K-means clustering algorithm, and the optimal number of clusters K is determined based on the silhouette coefficient. The geometric centroid of each cluster is solidified into a standard process morphology prototype. In the real-time processing flow, the feature extraction engine calculates the Euclidean distance between the newly generated fingerprint vector and each standard process morphology prototype, mapping... The vector is a unique discrete state code corresponding to the nearest prototype. The code is written as a node attribute into the heterogeneous manufacturing process graph, transforming the continuous floating-point feature space into a frequent subgraph for mining a finite set of discrete states. To maintain the consistency of the time-series difference operation dimension of the structured process fingerprint vector, the feature extraction process executes phase template forced alignment logic. The system pre-sets full-length feature templates with three standard time-series slots for heating and pressurization, constant temperature and pressure holding, and cooling and depressurization. When the first derivative sequence morphological segmentation anchor point judgment function is missing a sub-phase, the value of the local phase feature sub-vector is forcibly set to zero and filled into the corresponding slot of the template. The morphological segmentation logic change rate threshold is statistically analyzed based on the distribution of the absolute value of the first derivative of 200 gold sample waveforms. The 95th percentile is taken as the steady-state judgment boundary, and 50% of the average peak value of the derivative in the action segment is taken as the transient judgment boundary.

[0032] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0033] It should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application.

Claims

1. An industrial big data analysis method for micro-inverter manufacturing, characterized in that, The method comprises the following steps: parsing a discrete event log of a manufacturing execution system, taking a single product unique identification code as a core node, and constructing a directed traceability topology graph containing directed connection edges pointing to equipment nodes and material batch nodes; in response to process entry events and process exit events recorded in the discrete event log, defining a relative process time window of a single product in a specific process; obtaining equipment continuous time series data corresponding to the process, intercepting equipment continuous time series data within the relative process time window, and calculating a first derivative sequence of the equipment continuous time series data with respect to time; identifying time points at which the sign of the values in the first derivative sequence reverses or the values cross a preset change rate threshold, and defining the time points as morphological segmentation anchor points; logically dividing the equipment continuous time series data into multiple continuous functional sub-phases according to the morphological segmentation anchor points, and calculating local statistical features of each functional sub-phase; splicing the local statistical features of all functional sub-phases according to the time sequence logic to generate a structured process fingerprint vector, and writing the structured process fingerprint vector as an attribute value into the directed connection edge between the corresponding equipment node and the single product node in the directed traceability topology graph to generate a heterogeneous manufacturing process graph; obtaining single product final test failure data, performing reverse path traversal in the heterogeneous manufacturing process graph with the failure single product node as the terminal point, extracting a candidate path containing equipment nodes, material batch nodes and structured process fingerprint vectors, and mining subgraph structures with common feature attributes appearing more than a preset support threshold in the candidate path, and determining the subgraph structures as failure cause modes.

2. The industrial big data analysis method manufactured by the micro-inverter according to claim 1, wherein, The step of logically dividing the equipment continuous time series data into multiple continuous functional sub-phases according to the morphological segmentation anchor points comprises: locating a time point at which the first value greater than the positive change rate threshold in the first derivative sequence, defining the time point as a rising edge start anchor point; locating a time point at which the first value absolute value is less than the steady state change rate threshold and the duration exceeds a preset judgment time, defining the time point as a platform steady state anchor point; locating a time point at which the first value is less than the negative change rate threshold, defining the time point as a falling edge start anchor point; and based on the time sequence positions of the rising edge start anchor point, the platform steady state anchor point and the falling edge start anchor point, dividing the equipment continuous time series data into a temperature rising and pressure increasing phase, a constant temperature and pressure maintaining phase and a temperature decreasing and pressure releasing phase.

3. The industrial big data analysis method manufactured by the micro-inverter according to claim 1, wherein, The step of calculating the local statistical features of each functional sub-phase comprises: for each functional sub-phase, calculating the arithmetic mean, standard deviation and peak value of the data points within the phase as basic statistical quantities; calculating the duration proportion of the functional sub-phase within the relative process time window as a time domain feature quantity; and combining the basic statistical quantities and the time domain feature quantity to generate a local feature sub-vector of the functional sub-phase.

4. The industrial big data analysis method manufactured by the micro-inverter according to claim 3, wherein, The step of splicing the local statistical features of all functional sub-phases according to the time sequence to generate a structured process fingerprint vector comprises: connecting the local feature sub-vectors corresponding to each functional sub-phase in series according to the occurrence order of the functional sub-phases in the physical process; and performing normalization processing on the connected vector to generate a structured process fingerprint vector with fixed dimensions, so that the vector represents the topological evolution form of the device continuous time sequence data in the time dimension.

5. The industrial big data analysis method manufactured by the micro-inverter according to claim 1, wherein, The step of constructing the directed traceability topology graph further comprises constructing a time-adjacency chain for the same device node, specifically comprising: performing time sorting on all single-product nodes processed by the same device node according to the end time of the discrete event log; establishing a directed time-adjacency edge between the adjacent pre-sequence single-product node and the post-sequence single-product node; calculating a difference feature vector ΔV between the structured process fingerprint vector of the post-sequence single-product node and the structured process fingerprint vector of the pre-sequence single-product node, the calculation rule of the difference feature vector ΔV being: ΔV = V n -V n-1 , wherein V n is the structured process fingerprint vector of the post-sequence single-product node, and V n-1 is the structured process fingerprint vector of the pre-sequence single-product node; mapping the difference feature vector ΔV as an evolution attribute representing the dynamic drift of the device state to the directed time-adjacency edge.

6. The industrial big data analysis method manufactured by the micro-inverter according to claim 1, wherein, The step of mining sub-graph structures with high-frequency common feature attributes in the candidate paths comprises: inputting all extracted candidate paths into a frequent sub-graph mining algorithm; counting the occurrence frequency of each sub-graph structure in the candidate path set; identifying common sub-graph structures with an occurrence frequency exceeding a preset support threshold; and when a specific structured process fingerprint vector attribute interval is included in the common sub-graph structure, combining the specific attribute interval and the corresponding device node to lock the root cause mode leading to failure.

7. The industrial big data analysis method manufactured by the micro-inverter according to claim 1, wherein, The step of defining the relative process time window of a single product in a specific process further comprises performing logical alignment of multi-source data: when the sampling frequency of the device continuous time sequence data is higher than the time resolution of the discrete event log, taking the time when the process entry event occurs as a relative zero point; uniformly mapping the time axis of the device continuous time sequence data to an offset time axis relative to the relative zero point; and removing the data segments in the offset time axis after the occurrence time of the process exit event, and retaining the valid data segments in the closed interval of the process entry event and the process exit event. 8.The industrial big data analysis method manufactured by the micro-inverter according to claim 1, wherein, The step of writing the structured process fingerprint vector as an attribute value into the directed connection edge between the corresponding device node and the single product node in the directed root-seeking topological graph further comprises: obtaining the unique identification code of the auxiliary clamp used in the process and the identity code of the current operator as discrete attributes, and writing them into the directed connection edge between the device node and the single product node; and making the directed connection edge simultaneously carry continuous quantity features representing the physical process form and discrete quantity features representing the production resource configuration. 9.The industrial big data analysis method manufactured by the micro-inverter according to claim 6, wherein, The method further comprises the step of performing active blocking based on the mined failure cause mode: real-time monitoring of the heterogeneous manufacturing process graph generated in the subsequent production process; calculating the similarity between the connection attributes or sub-graph structures between the newly generated single product node and the device node and the failure cause mode; When the similarity exceeds a preset safety threshold, outputting a shutdown inspection instruction or a parameter correction instruction for the device.

Images