A filter-based high-end battery intelligent factory cloud-level process data analysis method

Abstract

This invention discloses a cloud-level process data analysis method for high-end battery smart factories based on filtering. It performs cloud-level aggregation and preprocessing of multi-source process data from various processes and end-level equipment in high-end battery production, constructing observation vectors, process input vectors, and cloud-level process state vectors. A discrete nonlinear state-space model is established, and the nonlinear state function and measurement function are linearized. Multi-cell filtering prediction and measurement updates are performed to obtain the multi-cell feasible region of the process state. Within the feasible region, partitioned gray wolf optimization and directional selective contraction are performed. Based on the contracted multi-cell set, the upper and lower bounds of the process state, the interval width, and anomaly risk information are calculated, and the cloud-level process data analysis results are output. This invention can achieve dynamic interval analysis of complex process states under unknown but bounded uncertainty conditions, improving the robustness, reliability, and interpretability of cloud-level process data analysis, and has good application prospects and promotional value.

Description

Technical Field

[0001] This invention belongs to the fields of high-end battery intelligent manufacturing, industrial big data analysis and intelligent information processing technology, specifically involving a cloud-level process data analysis method for high-end battery intelligent factories based on filtering. Background Technology

[0002] As high-end battery manufacturing continues to evolve towards digitalization, networking, and intelligence, the scale of process data in the production process is constantly expanding. The data sources are also gradually expanding from signals from single equipment to heterogeneous data sets covering multiple processes, equipment, and levels, including batching, coating, rolling, slitting, stacking, electrolyte injection, formation, and capacity testing. Various sensors, online monitoring devices, actuators, and manufacturing information systems continuously generate large amounts of time-series data, status data, and process quality data. This data is typically aggregated via industrial networks to a cloud platform for unified storage, management, and analysis. How to effectively analyze the multi-source process data of a high-end battery smart factory in a cloud environment, thereby accurately characterizing the process operating status, identifying process fluctuation trends, and supporting subsequent optimization decisions, has become a crucial problem that urgently needs to be solved in the field of intelligent manufacturing.

[0003] Current data processing methods in high-end battery production often employ process analysis approaches based on empirical rules, threshold judgments, statistical analysis, or conventional machine learning. While these methods can achieve data monitoring, anomaly identification, and quality assessment to some extent, they typically focus on static indicators or single-moment samples, making it difficult to fully characterize the temporal evolution characteristics, cross-process coupling relationships, and dynamic correlations between multiple data sources prevalent in high-end battery manufacturing. Furthermore, cloud-based process data often exhibits characteristics such as inconsistent sampling frequencies, different units, local missing data, and frequent abnormal disturbances. Relying solely on conventional statistical analysis or point-value prediction methods is insufficient to provide a continuous, stable, and reliable description of the real-world changes in complex production processes.

[0004] On the other hand, traditional Kalman filtering and its extensions have been widely used in industrial process state estimation and data processing. However, these methods generally rely on the assumption that process noise and measurement noise satisfy a known statistical distribution, especially a Gaussian distribution. In the actual production environment of high-end battery smart factories, factors such as process disturbances, equipment aging, sensor drift, network latency, data gaps, and modeling biases are prevalent, making it difficult to accurately obtain the statistical characteristics of process noise and measurement noise, and they may not even satisfy the assumption of a fixed probability distribution. Under these circumstances, traditional filtering methods based on probabilistic statistical models are prone to problems such as increased estimation bias, decreased stability, or insufficient reliability of results, making it difficult to meet the robustness and reliability requirements of cloud-level process data analysis.

[0005] To overcome the limitations imposed by stochastic statistical assumptions, set membership estimation and ensemble filtering methods have begun to be applied to state analysis and data processing scenarios under unknown but bounded uncertainty conditions. These methods, by maintaining a feasible set of system states in the prediction and update steps, can provide state interval or set estimation results when process disturbances and measurement noise only satisfy bounded constraints. Ellipsoidal sets, interval sets, and fully symmetric multicell sets, among other set representations, have been used to characterize the uncertainty of complex systems. However, existing ensemble filtering methods still have shortcomings when analyzing cloud-level process data from high-end battery smart factories: First, battery manufacturing processes exhibit significant nonlinearity, strong coupling, and multi-timescale characteristics; directly using simple linear models or conventional ensemble propagation methods can easily lead to inaccurate characterization of the process states. Second, multi-source process data, after being aggregated in the cloud, has high dimensionality and complex correlations; traditional ensemble wrapping and recursion processes can easily lead to increased conservatism and excessively wide feasible regions. Third, most existing methods focus on state estimation itself, with insufficient further exploration of high-probability regions within the set, making it difficult to effectively shrink the feasible region while ensuring feasibility, thus affecting the resolution and practical value of cloud-level analysis results.

[0006] Furthermore, existing process data analysis solutions typically output single predicted values ​​or single discrimination results, lacking a comprehensive characterization of process state upper and lower bounds, fluctuation width, confidence range, and risk trends. This makes it difficult to meet the "explainable, quantifiable, and traceable" analytical requirements of high-end battery smart factories in areas such as process monitoring, quality traceability, anomaly early warning, and cloud-based decision support. Therefore, there is an urgent need for a process data analysis method for cloud-based scenarios in high-end battery smart factories. This method should be able to establish corresponding dynamic analysis models for multi-process, multi-equipment, and multi-source heterogeneous process data, achieving reliable filtering analysis of process operating states under unknown but bounded uncertainty conditions. It should also reduce the conservatism of ensemble estimation while ensuring feasibility, thereby improving the accuracy, stability, and engineering application value of cloud-based process data analysis. Summary of the Invention

[0007] To address the following issues in existing cloud-based process data analysis for high-end battery smart factories: the difficulty in achieving unified, dynamic, and reliable state representation after the aggregation of heterogeneous process data from multiple processes, equipment, and sources in the cloud; the strong dependence of traditional statistically hypothetical filtering methods on noise distribution, making it difficult to adapt to the common unknown but bounded disturbances, sensor drift, data gaps, and modeling errors in actual production environments; the tendency of existing ensemble estimation methods to exhibit significant outsourcing conservatism in high-dimensional complex process scenarios, resulting in excessively wide analysis intervals and insufficient utilization of effective information; and the fact that existing methods typically only output single-value results, making it difficult to simultaneously provide interpretable analysis results such as upper and lower bounds of process states, fluctuation widths, and risk trends. Therefore, this invention provides a filtering-based cloud-based process data analysis method for high-end battery smart factories.

[0008] A cloud-based process data analysis method for high-end battery smart factories based on filtering includes the following steps: Step 1: Acquire process data generated by each process and end-level equipment during the high-end battery production process and upload it to the cloud analysis platform. Organize the aggregated process data to form observation vectors, process input vectors, and corresponding cloud-level process state vectors. Step 2: Based on process data, establish a discrete nonlinear state-space model to characterize the dynamic evolution of the high-end battery production process, and determine the process state vector, state transition function, measurement function, and unknown but bounded process disturbances and measurement noise; Step 3: Linearize the discrete nonlinear state-space model at the current state estimation result, construct the corresponding equivalent linear analysis model, and uniformly represent the linearization remainder, process disturbance and measurement noise as multi-cell set uncertainty; Step 4: Perform set prediction and measurement correction on the process state based on the equivalent linear analysis model to obtain the predicted cell set and the updated cell set of the process state, thereby forming a feasible region containing the actual process state. Step 5: Conduct a partitioned gray wolf search within the feasible region, and combine it with directional selective shrinkage to perform boundary compression on the updated multicellular set, so as to reduce the conservatism of the analysis while maintaining the feasibility of the set. Step 6: Based on the shrunken multi-cell set, determine the upper and lower bounds of the process state, the interval width, and the abnormal risk information, and generate the corresponding cloud-level process data analysis results.

[0009] Optionally, in step 1, the process data includes equipment operating parameters, process parameters, environmental parameters, and quality characterization parameters in the processes of batching, coating, rolling, slitting, stacking or winding, liquid injection, formation, capacity testing, aging, and testing. Step 1 also includes time synchronization, anomaly removal, missing data compensation, normalization, and feature alignment of the aggregated process data to form observation vectors. Process input vector and cloud-level process state vector .

[0010] Optionally, the discrete nonlinear state-space model established in step 2 can be represented as follows: ; in, Indicates the first The cloud-level process state vector at any given moment. Indicates the first The cloud-level process state vector at any given moment. For the input matrix, Indicates the first The process input vector at time step, Indicates the first The observation vector at time t, Represents the state transition function. Represents the measurement function. and These represent unknown but bounded process disturbances and measurement noise, respectively.

[0011] Optionally, in step 3, at the current estimated point State function and measurement function Performing first-order linearization (i.e., Taylor expansion to first-order terms) on each side yields the following results: ; in, ; and Let be the Jacobian matrices of the state function and the measurement function at the current estimation point, respectively. and These are the linearized higher-order remainder terms. Further, the higher-order remainder terms... and A bounded envelope is constructed using the upper bound of the Hessian interval. For the higher-order remainders of each component of the state function and the measurement function, respectively:

[0012] in, The state function is represented by the first The upper bound of the linearized remainder term of each component. Represents the measurement function. The upper bound of the linearized remainder term of each component. Preferably, the linearized remainder term of the state function is constructed as a set of state remainder term cells.

[0013] The linearized remainder of the measurement function is constructed as a set of multicells of measurement remainders.

[0014] in, This represents the dimension of the process state vector. Denotes the dimension of the observation vector. If the process disturbance and measurement noise respectively satisfy...

[0015] The combined state uncertainty set and the combined measurement uncertainty set are respectively expressed as:

[0016] in, This represents the Minkowski sum operation. Therefore, the linearized higher-order remainders, process disturbances, and measurement noise are uniformly enveloped using a polycellular set, resulting in an equivalent linear analysis model. ; in, . This represents the combined state uncertainty term, which consists of the linearized remainder of the state function and the process disturbance. This represents the combined measurement uncertainty term, which consists of the linearized remainder of the measurement function and measurement noise. Through the above processing, the original nonlinear process data analysis problem can be transformed into a linear set recursive problem under unknown but bounded uncertainty conditions, thus providing a model basis for the multicell filtering prediction and measurement update in step 4.

[0017] Optionally, in step 4, let the first... The set of process states at any given time is Set prediction is performed based on the equivalent linear analysis model, that is, using the fully symmetric multicellular structure in the set membership algorithm. Represents a set of states, based on the current time. The state variables are estimated for the next time step, thus obtaining... The fully symmetric multicell expression for the true state at time +1. express Predict the center of a fully symmetrical multicell at all times. This represents the generating matrix of the fully symmetric multicell; based on the linear state-space model obtained in step 3, for the... If we predict the state at a given time, then we have:

[0018] Thus, the predicted multicell set is obtained. ; in, ; in, and These are the generator matrices corresponding to the process perturbation set and the state remainder set, respectively; according to the... Constructing measurement strips from time-observed data ; in, and These are the generator matrices corresponding to the measurement noise set and the measurement remainder set, respectively; The predicted multicell set With measurement strip The intersection is then used to obtain the updated polytopic set through external approximation. ; in, ; To reduce the conservatism of the updated set, update gain Take as: .

[0019] Through the above prediction and update process, a fully symmetric multicell set containing the actual process state can be obtained. This fully symmetric multicell set maintains the inclusiveness of the actual state on the one hand, and shrinks the predicted feasible region by using measurement strip constraints on the other hand, thus forming the initial feasible search space for subsequent partitioned gray wolf optimization and direction-selective shrinking.

[0020] Optionally, in step 5, let the updated cell set obtained in step 4 be... ,in, To update the multicellular aggregation center, The generator matrix; Initialize the gray wolf population within the updated polycellular set, the first The initial positions of the individual gray wolves are represented as follows: ; And based on the generator matrix The directional information is used to partition the search space, thereby ensuring that the initial gray wolf individuals are all located within the updated polycellular feasible domain.

[0021] Furthermore, based on the generator matrix The direction information is used to partition the search space. Let the direction information selected from the generator be used to partition the search space. The normalized principal direction vectors are denoted as follows: For any iteration time Gray Wolf's position The displacement vector relative to the center of the set is defined as ; Its partition satisfies ; in, This is the selected principal direction vector.

[0022] This divides the updated multicellular feasible region into several partitions (or "sub-regions"). Optionally, within each partition, the three gray wolf individuals with the best fitness in that region are selected as the optimal groups. Wolf, wolves and Wolves, respectively denoted as , and And the globally optimal individual is denoted as ; by the region Wolf, wolves and The wolf constructs a local guiding term, which is composed of the globally optimal individual and region. The wolf constructs global guiding terms, which are represented as follows: ; Correspondingly, the location of the gray wolf has been updated to ; in, ; The fitness function is determined based on the deviation between the predicted output and the actual observation, and can be expressed as follows:

[0023] Furthermore, to ensure that the iterated gray wolf individuals always remain within the original updated polycellular feasible region, an out-of-bounds criterion is set for the updated gray wolf individuals. Let...

[0024] Then when satisfied When the gray wolf individual is located in a multicellular feasible region, it is determined that the individual gray wolf is located in the multicellular feasible region. If the boundary is not within the specified range, then perform boundary correction on the out-of-bounds individuals to bring them back into the said multicellular feasible domain.

[0025] Optionally, in step 5, after completing the gray wolf iterative optimization, a principal direction analysis is performed based on the optimized sample distribution. Let the obtained eigenvector matrix be... For each main direction Calculate the projection range of the optimized sample along this direction. ; And define the target half-width in the corresponding direction as ; in, To allow for a safety margin, the projected radius in the corresponding direction is then calculated based on the polycellular support function. , to obtain the contraction factor The contraction matrix is ​​constructed from the contraction factors in each direction. ; Then, the original generator matrix is ​​directionally selectively shrunk to obtain... Thus, the shrunken multicell set is obtained. .

[0026] Directional selective shrinkage performs principal direction analysis on the optimized sample distribution to determine the target half-width in different principal directions, and calculates the shrinkage factor in the corresponding direction based on the cell support function, thereby constructing a shrinkage matrix to non-uniformly scale the original cell set and reduce redundant outlay in the traditional set propagation process.

[0027] Optionally, in step 6, based on the shrunken multicell set ,in, Indicates the multicell aggregation center after contraction. This represents the shrunk generator matrix.

[0028] Calculate the upper and lower bounds and interval width of the process state; for any process state component Its lower boundary and the Upper Realm They are respectively ; in, Represents the center vector The One portion, Represents the generator matrix The Middle Line number Column elements. Therefore, the corresponding interval width is... ; And with the central estimate This serves as the central analysis result of the current cloud-level process status.

[0029] Based on the upper and lower bounds of the process state, the interval width, and the preset reference state, cloud-level process deviation and anomaly risk indicators are constructed. Let the reference state vector be... The process deviation is then expressed as ; make ; The abnormal risk indicator is expressed as: ; in, It is a vector composed of the widths of the intervals of each state component. and These are the upper bounds of the deviation term and the interval width term, respectively. and For the preset weighting coefficients, satisfy and These are used to characterize the influence of the process state deviation term and the interval uncertainty term on the anomaly risk index, respectively. Preferably, when the process analysis focuses more on the degree of deviation from the reference operating condition, a larger value is taken. When process analysis focuses more on the range of state fluctuations and the level of uncertainty, the larger value should be selected. .

[0030] When the abnormal risk indicator Greater than the preset threshold If an abnormal risk is detected, the system determines that the current process status is within the normal analysis range and outputs the corresponding cloud-level process data analysis results. Otherwise, the system determines that the current process status is within the normal analysis range. Based on the upper and lower bounds of the process status, the interval width, the deviation, and the abnormal risk indicators, corresponding cloud-level process data analysis results are generated for process monitoring, quality traceability, process optimization, or abnormal early warning.

[0031] The beneficial effects of this invention are as follows: This invention addresses the dynamic analysis needs of multi-process, multi-equipment, and multi-source heterogeneous process data in cloud-based scenarios of high-end battery smart factories. By constructing a nonlinear state-space model, introducing unknown but bounded uncertainty descriptions, combining a multi-cell filtering prediction update mechanism, and employing partitioned gray wolf optimization and directional selective contraction strategies, it achieves interval-based, dynamic, and interpretable analysis of complex process states. Compared to existing analysis methods that rely on fixed statistical assumptions or only output single-value results, this invention effectively reduces conservatism in the set propagation process while ensuring state feasibility inclusion, improving the robustness, reliability, and discriminative power of cloud-based process data analysis, and providing more comprehensive analytical basis for process monitoring, quality traceability, anomaly early warning, and process optimization. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the overall process of a cloud-based process data analysis method for a high-end battery smart factory based on filtering, as described in an embodiment of the present invention.

[0033] Figure 2 This is a schematic diagram of the cloud-based process data analysis architecture for a high-end battery smart factory, as described in an embodiment of the present invention.

[0034] Figure 3 This is a graph showing the cloud-level process state ×1 interval estimation results of the cloud-level process data analysis method for high-end battery smart factories based on filtering in this embodiment of the invention.

[0035] Figure 4 This is a graph showing the cloud-level process state ×2 interval estimation results of the cloud-level process data analysis method for high-end battery smart factories based on filtering in this embodiment of the invention.

[0036] Figure 5 This is a graph showing the cloud-level process state ×3 interval estimation results of the cloud-level process data analysis method for high-end battery smart factories based on filtering in this embodiment of the invention.

[0037] Figure 6 This is a diagram of abnormal risk indicators under normal operating conditions for the cloud-level process data analysis method for high-end battery smart factories based on filtering, as described in this embodiment of the invention.

[0038] Figure 7This is an abnormal risk index diagram under abnormal operating conditions in the middle section of the cloud-level process data analysis method for high-end battery intelligent factories based on filtering, as described in this embodiment of the invention. Detailed Implementation

[0039] Example 1 This embodiment provides a cloud-based process data analysis method for high-end battery smart factories based on filtering. The process flow of this method is as follows: Figure 1 As shown. The method is executed in a cloud-based analysis platform and includes six stages: cloud-level process data aggregation (acquiring high-end battery production process data and constructing observation, input, and state vectors), cloud-level state modeling (establishing a discrete nonlinear state-space model), equivalent linear set model construction (linearizing the nonlinear model and using fully symmetric multicells to wrap linearization errors and noise uncertainties), multicell filtering prediction and updating (predicting and updating based on the state-space model to obtain the initial fully symmetric multicell feasible set search space), partitioned gray wolf optimization and directional selective contraction (conducting partitioned gray wolf search within the feasible region and combining directional selective contraction to compress multicell boundaries), and analysis result output (calculating the upper and lower bounds of the state, interval width, and anomaly risk to obtain cloud-level process data analysis results).

[0040] Step 1 involves cloud-based aggregation of process data from various production processes and end-level equipment in the high-end battery smart factory. Specifically, data from different processes, equipment, and sampling frequencies are uniformly uploaded to the cloud platform, and time synchronization, anomaly removal, missing data compensation, normalization, and feature alignment are sequentially performed to form a unified data input. Based on the preprocessing results, the first... Time observation vector Process input vector And the cloud-level process state vector to be analyzed. .

[0041] Among them, the cloud-level process state vector to be analyzed This is used to characterize implicit state information in high-end battery production processes that is difficult to measure directly and completely, but can reflect the process operation status, quality evolution trend, and multi-process coupling relationship; the process input vector The observation vector is used to characterize control variables, setpoints, or process condition parameters that drive the evolution of process states; Used to characterize measurable process data uploaded to a cloud platform by sensors, online inspection devices, edge acquisition units, or manufacturing execution systems.

[0042] Step 2: After completing the data construction, establish a discrete nonlinear state-space model:

[0043] in, Indicates the first The cloud-level process state vector at any given moment. Indicates the first The cloud-level process state vector at any given moment. For the input matrix, Indicates the first The process input vector at time step, Indicates the first The observation vector at time t, Represents the state transition function. Let B represent the measurement function, and let B represent the input matrix. and These represent unknown but bounded process disturbances and measurement noise, respectively. The process disturbances are mainly used to characterize state evolution deviations caused by factors such as equipment fluctuations, changes in operating conditions, environmental disturbances, time-varying parameters, and model mismatches; the measurement noise is mainly used to characterize observation deviations caused by factors such as sensor drift, detection errors, sampling jitter, and abnormal data uploads.

[0044] Step 3, in this embodiment, no prior assumptions are made. and It follows a fixed probability distribution, but is only considered to satisfy unknown but bounded constraints. Accordingly, a fully symmetric polytopic set is used to uniformly describe the set of states and uncertainties. Any polytopic set... Represented as:

[0045] in, Indicates the center of the set. Represents the generator matrix, This represents a coefficient vector that satisfies an infinity norm of no more than 1.

[0046] At the current estimation point At this point, the nonlinear state function and the measurement function are respectively linearized to the first order, resulting in:

[0047] in,

[0048] and Let represent the Jacobian matrices of the state function and the measurement function at the current estimation point, respectively; and These represent the higher-order remainder terms generated by linearization; The symbol represents the partial derivative.

[0049] To facilitate subsequent set recursive analysis, the constant terms and higher-order remainder terms generated by linearization are uniformly incorporated into the uncertain terms, resulting in the following equivalent linear analysis model:

[0050] in, This represents the overall state uncertainty. This indicates the uncertainty term in the overall measurement.

[0051] Furthermore, using the upper bound of the Hessian interval to... and Perform a bounded envelope. For each component of the state function and the measurement function, we have:

[0052] in, Indicates time state function vector The One portion, Indicates time Measurement function vector The One component; and These represent the upper bounds of the absolute values ​​of the higher-order remainders of the corresponding components after linearization; , ,in Let be the dimension of the process state vector. Let be the dimension of the observation vector.

[0053] Based on this, construct the state remainder polycell set and the measurement remainder polycell set, denoted as follows: and If the process disturbance and measurement noise satisfy The combined state uncertainty set and the combined measurement uncertainty set are respectively expressed as:

[0054] in, This represents Minkowski operations.

[0055] Step 4, let the first... k The time-state estimation set is as follows Then, based on the equivalent linear set model, a prediction is performed to obtain the first... Predictive multicell set at time:

[0056] in,

[0057] in, and These represent the generator matrices corresponding to the process perturbation set and the state remainder set, respectively.

[0058] Based on the The observation data at each time point are used to construct measurement strips:

[0059] in, and These represent the generator matrices corresponding to the measurement noise set and the measurement remainder set, respectively. The predicted cell set... With the measurement strip Intersect and use an external approximation form to obtain the updated polytopic set:

[0060] in,

[0061] The update gain Take as:

[0062] I This is an identity matrix that matches the dimension of the process state vector. Through the above prediction and update process, an updated multicellular feasible region containing the actual process state can be obtained.

[0063] Step 5, in the obtained updated multicellular feasible domain Afterwards, among them, To update the multicellular aggregation center, The generator matrix; Initialize the gray wolf population within the feasible region. Let the first... The initial position of each gray wolf for:

[0064] in, This ensures that all initial gray wolf individuals are located within the updated multicellular feasible domain.

[0065] Based on the generator matrix The directional information is used to partition the search space. Let the selected... The normalized principal direction vectors are denoted as follows: For any iteration time Gray Wolf's position The displacement vector relative to the center of the set is defined as Then its partition for:

[0066] This divides the updated multicellular feasible region into several partitions.

[0067] Within each partition, the optimal region is selected based on its fitness. Wolf, wolves and Wolves, respectively denoted as , and And the globally optimal individual is denoted as: This leads to the construction of local guiding terms. and global bootstrap :

[0068] Correspondingly, the gray wolf's location has been updated to:

[0069] in,

[0070] The fitness function is determined based on the deviation between the predicted output and the actual observation. Simultaneously, to ensure that each updated gray wolf individual remains within the original updated multicellular feasible region, an out-of-bounds criterion is set for each updated individual. Let...

[0071] Then when satisfied

[0072] hour, To update the multicell set generator matrix of the first... Column-generated vectors To generate vector indices, , The column number of the generator matrix; determine if the gray wolf individual is located in the multicellular feasible region. If the boundary is not within the specified range, then perform boundary correction on the out-of-bounds individuals to bring them back into the said multicellular feasible domain.

[0073] After completing the gray wolf iterative optimization, principal direction analysis is performed based on the optimized sample distribution. Let the eigenvector matrix obtained from the principal direction analysis be... For each main direction Calculate the projection range of the optimized sample along this direction:

[0074] Define the target half-width in the corresponding direction. for:

[0075] in, For safety margin, the projected radius in this direction is then calculated based on the polycellular support function. And obtain the contraction factor. The contraction matrix is ​​constructed from the contraction factors in each direction.

[0076] Then, the original generator matrix is ​​directionally selectively shrunk to obtain...

[0077] This results in a condensed multicell set.

[0078] Step 6: Calculate the upper and lower bounds and interval width of the process state based on the shrunken multicell set obtained in Step 5. For any process state component... Its lower boundary and the Upper Realm They are respectively represented as

[0079] in, This represents the generator matrix after shrinkage. The Middle Line number Column elements; q Represents the generator matrix after shrinkage The number of columns, i.e. the number of generated vectors; This is the index for the process state component. The interval width corresponding to the process state component. for

[0080] and with This serves as the central analysis result of the cloud-level process status at the current moment. Furthermore, an anomaly risk index is constructed based on the process status center deviation and interval width. Let the reference state vector be... Then the process deviation amount It can be represented as

[0081] make Then abnormal risk indicators It can be represented as

[0082] in, and For the weighting coefficients, satisfying and When the process analysis focuses more on the degree to which the operating condition deviates from the reference condition, a larger value should be taken. When process analysis focuses more on the range of state fluctuations and the level of uncertainty, the larger value should be selected. .

[0083] When the abnormal risk indicator Greater than the preset threshold If an abnormal risk is detected, the current process status is determined to be within the normal analysis range; otherwise, it is determined to be within the normal analysis range. Based on the upper and lower bounds of the process status, the interval width, the deviation, and the abnormal risk indicators, corresponding cloud-level process data analysis results are generated for process monitoring, quality traceability, process optimization, or abnormal early warning.

[0084] In this embodiment, the multi-source process data analysis problem is transformed into a dynamic set recursion problem under unknown but bounded uncertainty conditions. Furthermore, based on multi-cell filtering, partitioned gray wolf optimization and directional selective contraction are introduced to achieve interval-based, dynamic, and interpretable analysis of process states. This method can reduce the conservatism of set propagation while maintaining inclusiveness of the actual process states, thereby improving the effectiveness and reliability of cloud-level process data analysis.

[0085] Example 2 This embodiment provides a cloud-based process data analysis architecture for a high-end battery smart factory based on filtering. For example... Figure 2 As shown, the architecture includes a data acquisition layer, an edge aggregation layer, a cloud analysis layer, and a result application layer. The data acquisition layer is used to collect process data for each step in the high-end battery production process. The edge aggregation layer is used to perform data caching, preliminary cleaning, protocol conversion, and upload control. The cloud analysis layer is used to execute the cloud-level process data analysis method in Example 1. The result application layer is used to receive analysis results and serve process monitoring, quality traceability, anomaly warning, and process optimization.

[0086] Specifically, the data acquisition layer can be deployed in processes such as batching, coating, rolling, slitting, stacking or winding, liquid injection, formation, capacity testing, aging, and testing. The corresponding data acquisition objects include equipment operating parameters, process parameters, environmental parameters, and quality characterization parameters. The acquired data is uniformly encapsulated by the edge acquisition unit and then uploaded to the cloud analysis platform via industrial Ethernet, fieldbus, or wireless industrial communication networks. To ensure the comparability of data from different sources in the cloud, the edge aggregation layer includes not only basic data caching and upload control but also edge acquisition units, a data pre-cleaning module, and a data caching module. The edge acquisition unit is used for multi-source data access, protocol conversion, and local caching; the data pre-cleaning module performs timestamp alignment, outlier removal, and missing value marking; the data caching module performs batch organization, cache management, and breakpoint resumption; and the upload control module implements link management, encrypted transmission, and upload control. Through the above processing, basic standardization and reliable transmission can be completed before the data enters the cloud.

[0087] The cloud-based analysis layer includes at least a data preprocessing module, a state modeling module, an ensemble prediction and update module, an optimization and shrinking module, and a result output module. The data preprocessing module receives the data stream uploaded from the edge aggregation layer and further performs cross-process data association, sampling frequency unification, and feature alignment to construct the observation vector, process input vector, and cloud-level process state vector to be analyzed, as described in Example 1. The state modeling module establishes a discrete nonlinear state-space model and calculates the Jacobian matrix online based on the current state estimation results, while simultaneously performing linearization of the remainder term, bounded process perturbation, and the ensemble envelope of measurement noise.

[0088] The set prediction and update module is used to propagate the process state set in time according to the prediction-update recursive method in Example 1. Specifically, at each analysis time, the set prediction and update module reads the updated cell set from the previous time, combines it with the current process input and the equivalent linear set model to obtain the predicted cell set; then, based on the current observation data, it constructs measurement strips and performs intersectional out-approximation update on the predicted cell set to obtain the updated cell set. Through the above method, the cloud analysis layer can continuously output the dynamic feasible region corresponding to the process operation state.

[0089] The optimization shrinkage module is used to further reduce the conservatism of the set outer envelope based on the updated multicell set. Specifically, the optimization shrinkage module first initializes the gray wolf population within the updated multicell feasible region and divides the search partition according to the direction information of the generator matrix; then, it constructs local and global guiding terms using regional and global guiding individuals to iteratively update the position of the gray wolf population; after completing the iteration, it performs principal direction analysis on the optimized samples, constructs a shrinkage matrix, and generates the shrunken multicell set. The result output module calculates the upper bound, lower bound, and interval width of each process state component based on the shrunken multicell set, and further combines the reference state vector to construct deviation and anomaly risk indicators, forming cloud-level process data analysis results.

[0090] To further illustrate the analysis output format under the architecture of this invention, this embodiment selects three typical cloud-level process state components as examples in the cloud analysis layer, denoted as follows: , and By performing data analysis on the state components based on the method described in Example 1, the estimated state center value, upper bound of the state, lower bound of the state, and corresponding interval width can be obtained at each time point.

[0091] Figures 3 to 5 Representing state components respectively , and The interval estimation results. From Figures 3 to 5 It can be seen that the actual evolution trajectories of each state component are all located within the intervals formed by their corresponding upper and lower bounds, indicating that the present invention can effectively encompass the actual process state under unknown but bounded uncertainty conditions. Simultaneously, the center estimation results can well characterize the state change trend, demonstrating that the present invention can achieve dynamic interval analysis of multidimensional cloud-level process states. Furthermore, by combining multi-cell filtering prediction updates with partitioned gray wolf optimization and directional selective contraction, the present invention can effectively reduce redundant outsourcing in the set propagation process while maintaining state feasibility inclusion, thereby improving the resolution and interpretability of the process analysis results.

[0092] To illustrate the application effect of this invention in abnormal risk identification, two analysis scenarios are further constructed: normal operating conditions and intermediate abnormal operating conditions. Abnormal risk indicators are calculated based on the process state deviation and interval width. Figure 6 The results of the abnormal risk indicators under normal operating conditions are shown. Under normal operating conditions, each state component fluctuates within a bounded range around the reference operating condition, and the deviation and interval width remain within a small range. Therefore, the corresponding abnormal risk indicators are generally lower than the preset threshold, indicating that the current process is within the normal analysis range. Figure 7The results of the abnormal risk index under abnormal operating conditions in the middle stage are shown. In this scenario, the middle period of the process is affected by abnormal disturbances, which causes the state deviation term and the interval width term to increase simultaneously. This results in the abnormal risk index rising significantly during the corresponding period and exceeding the preset threshold. Therefore, it can be determined that there is an abnormal risk in the current process state, and the corresponding abnormal warning result is output.

[0093] In this embodiment, the method described in Example 1 can be implemented in a cloud-based analysis platform by software programs, firmware logic, or a combination of hardware and software. The cloud-based analysis platform can be deployed in a private cloud, industrial edge cloud, or hybrid cloud environment, and can operate for single-factory production lines or centralized analysis scenarios involving multiple production lines, workshops, or factories. Through the above-described architecture deployment, this invention can uniformly incorporate multi-source heterogeneous process data from high-end battery smart factories into a dynamic ensemble analysis framework, achieving range-based, dynamic, and interpretable analysis of complex processes while ensuring state inclusiveness.

Claims

1. A cloud-based process data analysis method for high-end battery smart factories based on filtering, characterized in that, Includes the following steps: Step 1: Acquire process data generated by each process and end-level equipment during the high-end battery production process and upload it to the cloud analysis platform. Organize the aggregated process data to form observation vectors, process input vectors, and corresponding cloud-level process state vectors. Step 2: Based on process data, establish a discrete nonlinear state-space model to characterize the dynamic evolution of the high-end battery production process, and determine the process state vector, state transition function, measurement function, and unknown but bounded process disturbances and measurement noise; Step 3: Linearize the discrete nonlinear state-space model at the current state estimation result, construct the corresponding equivalent linear analysis model, and uniformly represent the linearization remainder, process disturbance and measurement noise as multi-cell set uncertainty; Step 4: Perform set prediction and measurement correction on the process state based on the equivalent linear analysis model to obtain the predicted cell set and the updated cell set of the process state, thereby forming a feasible region containing the actual process state. Step 5: Conduct a partitioned gray wolf search within the feasible region, and combine it with directional selective shrinkage to perform boundary compression on the updated multicellular set, so as to reduce the conservatism of the analysis while maintaining the feasibility of the set. Step 6: Based on the shrunken multi-cell set, determine the upper and lower bounds of the process state, the interval width, and the abnormal risk information, and generate the corresponding cloud-level process data analysis results.

2. The cloud-based process data analysis method for high-end battery smart factories based on filtering according to claim 1, characterized in that, In step 1, the process data includes equipment operating parameters, process parameters, environmental parameters, and quality characterization parameters in the processes of batching, coating, rolling, slitting, stacking or winding, liquid injection, formation, capacity testing, aging, and testing. Step 1 also includes time synchronization, anomaly removal, missing data compensation, normalization, and feature alignment of the aggregated process data to form observation vectors. Process input vector and cloud-level process state vector .

3. The cloud-based process data analysis method for high-end battery smart factories based on filtering according to claim 1, characterized in that, The discrete nonlinear state-space model established in step 2 is expressed as follows: ； in, Indicates the first The cloud-level process state vector at any given moment. Indicates the first The cloud-level process state vector at any given moment. For the input matrix, Indicates the first The process input vector at time step, Indicates the first The observation vector at time t, Represents the state transition function. Represents the measurement function. and These represent unknown but bounded process disturbances and measurement noise, respectively.

4. The cloud-level process data analysis method for high-end battery smart factories based on filtering according to claim 3, characterized in that, In step 3, at the current estimated point State function and measurement function Perform first-order linearization on each, and obtain ； in, ; and Let be the Jacobian matrices of the state function and the measurement function at the current estimation point, respectively. and To linearize the higher-order remainders, the linearized higher-order remainders, process disturbances, and measurement noise are all enveloped using a multi-cell set to obtain an equivalent linear analysis model. ； in, This represents the overall state uncertainty. This represents the overall measurement uncertainty. .

5. The cloud-based process data analysis method for high-end battery smart factories based on filtering according to claim 4, characterized in that, In step 4, let the first... The set of process states at any given time is Based on the equivalent linear analysis model, set prediction is performed to obtain the predicted multicell set. ； in, ; in, and These are the generator matrices corresponding to the process perturbation set and the state remainder set, respectively; according to the... Constructing measurement strips from time-observed data ； in, and These are the generator matrices corresponding to the measurement noise set and the measurement remainder set, respectively; The predicted multicell set With measurement strip The intersection is then used to obtain the updated polytopic set through external approximation. ; in, ; Update gain Take as: .

6. The cloud-based process data analysis method for high-end battery smart factories based on filtering according to claim 5, characterized in that, In step 5, let the updated cell set obtained in step 4 be... ,in, To update the multicellular aggregation center, The generator matrix; Initialize the gray wolf population within the updated polycellular set, the first The initial positions of the individual gray wolves are represented as follows: ; And based on the generator matrix The directional information is used to partition the search space; for any gray wolf position The displacement vector relative to the center of the set is defined as ； Its partition satisfies ; in, This is the selected principal direction vector.

7. The cloud-based process data analysis method for high-end battery smart factories based on filtering according to claim 6, characterized in that, Within each partition, the three gray wolf individuals with the best fitness in that region are selected as the optimal groups. Wolf, wolves and Wolves, respectively denoted as , and And the globally optimal individual is denoted as ; by region Wolf, wolves and The wolf constructs a local guiding term, which is composed of the globally optimal individual and region. The wolf constructs global guiding terms, which are represented as follows: ； Correspondingly, the location of the gray wolf has been updated to ; in, ; An out-of-bounds criterion is set for the updated gray wolf individuals. If an individual exceeds the updated polycell set, boundary correction is performed to bring it back into the updated polycell set.

8. The cloud-level process data analysis method for high-end battery smart factories based on filtering according to claim 7, characterized in that, In step 5, after completing the gray wolf iterative optimization, principal direction analysis is performed based on the optimized sample distribution. Let the obtained eigenvector matrix be... For each main direction Calculate the projection range of the optimized sample along this direction. ; And define the target half-width in the corresponding direction as ; in, To allow for a safety margin, the projected radius in the corresponding direction is then calculated based on the polycellular support function. , to obtain the contraction factor The contraction matrix is ​​constructed from the contraction factors in each direction. ; Then, the original generator matrix is ​​directionally selectively shrunk to obtain... Thus, the shrunken multicell set is obtained. .

9. The cloud-based process data analysis method for high-end battery smart factories based on filtering according to claim 8, characterized in that, In step 6, based on the shrunken multicell set Calculate the upper and lower bounds and interval width of the process state; For any process state component Its lower boundary and the Upper Realm They are respectively ； The corresponding interval width is ; and with This serves as the central analysis result of the current cloud-level process status. Let the reference state vector be The process deviation is then expressed as ; make ; The abnormal risk indicator is expressed as: ; in, and These are the upper bounds of the deviation term and the interval width term, respectively. and For the preset weighting coefficients, satisfy and ; When the abnormal risk indicator Greater than the preset threshold When an abnormal risk is detected in the current process status, the corresponding cloud-level process data analysis results are output.

Images