Method for monitoring and early warning abnormal working conditions in precision forming process of cartridge case

Abstract

The invention provides a cartridge case precision forming process abnormal working condition monitoring and early warning method. Diffusion distance fuzzy clustering (DDFCM) and dynamic internal principal component analysis (DiPCA) technologies are fused. The method comprises the following steps: firstly, performing fuzzy clustering on process data based on a diffusion distance by adopting a DDFCM method, automatically identifying key sample groups under different working conditions, and constructing a representative local training data set; thirdly, for each clustering sample group, constructing a dynamic monitoring model by using a DiPCA technology, effectively extracting potential variables containing time lag features, and capturing a dynamic change mode of a process; in a new data acquisition stage, firstly, the membership degree of a new sample in each cluster is calculated through DDFCM, and the new sample is distributed to a corresponding local DiPCA model for state discrimination according to a maximum membership degree principle. According to the method, the time sequence dynamic change in the process can be fully captured, the local characteristics in different production states are also considered, and a powerful guarantee is provided for the production stability and the product quality of the cartridge case coating, rolling and cutting process.

Description

Technical Field

[0001] The present invention relates to the field of abnormal working condition monitoring in industrial processes, and in particular to a method for monitoring abnormal working conditions in precision molding processes such as coating of cartridges with energetic adhesives, rolling into cartridges, and chip molding. Background Art

[0002] Cartridges, with their simple and low-cost production process and dual functions as both container and energy source, are widely used as propellant charge containers for new artillery. Their manufacturing process involves precision machining of energetic materials and complex process control, placing extremely stringent safety requirements on production. During the cartridge forming process, the nitrocellulose paper undergoes critical steps such as energetic adhesive coating, high-temperature roll forming, and precision cutting. These steps require real-time control of high-risk process parameters such as coating tank temperature, paper feed tension, and coiling speed. Due to the inherent flammability and explosiveness of nitrocellulose and energetic adhesives, as well as the high-temperature and high-pressure production environment, any gradual deviation in process parameters or abnormal operating conditions can trigger a chain reaction. For example, excessive fluctuations in coating tank temperature can cause localized adhesive curing failure or overreaction, increasing the risk of uncontrolled combustion. Abnormal paper feed tension can directly disrupt the uniformity of the cartridge structure, leading to stress concentration during subsequent loading or firing, leading to rupture or even explosion. Furthermore, deviations in the dynamic matching of rolling speed, temperature, and tension can cause microcracks within the barrel or uneven distribution of energetic materials, potentially leading to serious safety accidents such as barrel explosion due to the impact of high temperature and high pressure at the moment of gun firing. Therefore, real-time monitoring and early warning of gradual deviations in process parameters and abnormal operating conditions are key prerequisites for preventing uncontrolled reactions of energetic materials and ensuring production safety.

[0003] Traditional anomaly monitoring methods (such as static modeling based on principal component analysis) have significant limitations: First, static models struggle to capture the lag effects and dynamic correlations of process parameters, resulting in insufficient sensitivity to early, subtle anomalies. Second, a single global model cannot adapt to sudden changes in data distribution under different operating conditions, and is prone to false positives or missed detections due to operating condition switching. Third, traditional clustering methods ignore the similarity of the global data structure, making it difficult to accurately segment multimodal operating condition subsets, which affects the characterization capabilities of local models. For example, when the coating tank temperature fluctuates due to environmental disturbances, the dynamic coupling relationship between the coil tension and paper feed speed may deviate from the normal mode. However, traditional methods lack the ability to extract dynamic features, making it difficult to promptly identify such hidden anomalies. Furthermore, existing technologies often rely on fixed thresholds or manual experience to set monitoring rules, making them difficult to cope with highly dynamic, highly coupled, and complex production processes. Therefore, there is an urgent need for an anomaly monitoring method that can both reflect the inherent temporal dynamics of the process and adapt to the characteristics of multiple operating conditions to improve monitoring sensitivity and real-time response capabilities. Summary of the Invention

[0004] To address the aforementioned issues in the prior art, the present invention proposes a method for monitoring and early warning abnormal conditions during the precision molding process of cartridges. This method primarily combines two advanced data processing techniques: diffusion distance-based fuzzy clustering (DDFCM) and dynamic internal principal component analysis (DiPCA). The specific scheme involves preprocessing the process data collected during the cartridge production process (such as coating tank temperature, rotation speed, paper feed speed, tension, and tube winding speed) to form a training dataset. Fuzzy clustering is then performed on the training data using the DDFCM method. Global data similarity is measured using diffusion distances to automatically divide sample groups under different operating conditions. A DiPCA dynamic monitoring model is then constructed for each clustered sub-dataset, extracting dynamic latent variables with time lag characteristics to capture the temporal variation patterns of the process during operation. Finally, during the new data collection phase, new samples are assigned to the corresponding local models based on the maximum membership degree. Real-time monitoring of abnormal conditions is achieved through the calculation of dynamic statistics and control limits. This method not only takes into account the dynamic evolution of the overall process but also accurately identifies local anomalies, effectively improving the real-time and accuracy of anomaly monitoring.

[0005] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:

[0006] A method for monitoring and early warning abnormal working conditions in a precision molding process of a cartridge, comprising the following steps:

[0007] 1) Collect data from the cartridge production process and normalize it to obtain the training data set D′ train ;

[0008] 2) Use the diffusion distance-based fuzzy clustering method to divide the training data set into stages and obtain M cluster sub-datasets D′ train =[D′ train1 ,D′ train2 ,…,D′ trainM ];

[0009] 3) A dynamic internal principal component analysis model is constructed for each cluster sub-dataset, and its objective function is optimized to extract dynamic latent variables;

[0010] 4) Calculate the core statistics of process monitoring based on dynamic latent variables and Q i , and obtain the control limit T 2 and Q;

[0011] 5) Repeat steps 3) and 4) until the dynamic internal principal component analysis model is constructed for all M cluster sub-datasets and the corresponding control limits are calculated;

[0012] 6) Calculate the new sample x tThe membership degree of the new sample x is analyzed using the dynamic internal principal component analysis model corresponding to the membership degree. t Conduct abnormal monitoring.

[0013] The step 2) comprises the following steps:

[0014] 2.1) Calculate the diffusion distance D(x) between samples i ,x j ):

[0015]

[0016] Among them, variables i, j, l = 1, 2...n, φ0(x) is the penalty difference factor corresponding to different areas of sample density, p2(x i ,x j ) is point x i To point x j The 2-step transition probability, deg(x l ) is the sample local density weight;

[0017] 2.2) Obtaining the degree of data membership in the global structure based on the diffusion distance fuzzy clustering method:

[0018]

[0019] Among them, J m is the objective function of the clustering algorithm, and m is the membership degree U pq The weight factor, is the 2-step transfer probability from point p to point q, i.e., the diffusion distance. When the objective function is minimized, the membership function U pq and cluster center c q for:

[0020]

[0021]

[0022] 2.3) Classify samples with membership within the same threshold range into the same fuzzy class.

[0023] The step 3) comprises the following steps:

[0024] 3.1) Define dynamic latent variable extraction rules:

[0025] For a time series vector x t , generate dynamic latent variables d(t)=w by projecting vector w T x t , whose predicted values ​​satisfy the multi-order autoregressive constraints: where β=[β1,...,β s ]T is the weight coefficient vector, s is the preset lag order;

[0026] 3.2) Construct dynamic optimization objective function:

[0027]

[0028] st||w||=1,||β||=1

[0029] in, represents the Kronecker product, where:

[0030] Z s =[X1,X2,...,X s ]

[0031] X i =[x i ,x i+1 ,...,x(N+i-1)] T

[0032] Among them, X i represents the data submatrix intercepted at the i-th moment in the time series, Z s Indicates multiple Xs i The collection after horizontal merger;

[0033] 3.3) After extracting the dynamic latent variable feature d = Xw, iteratively update X:

[0034] X:=X-dp T

[0035] Where p = X T d / (d T d) is the load vector;

[0036] 3.4) After step 3.3), multiple dynamic latent variables are extracted, and the corresponding load matrix P and feature extraction matrix W are obtained.

[0037] The step 4) is specifically as follows:

[0038] d r (t) = P r T e(t)

[0039] e r (t)=(IP r P r T )e(t)

[0040]

[0041] Among them, P ris the principal component projection matrix, I is the identity matrix, d r (t) is the static principal component, e r (t) is the static residual, e(t) is the static component, T 2 (t) is a statistic that measures the degree of deviation of the sample in the principal component subspace, I is the unit matrix, Q(t) represents the Q statistic calculated at time t, and finally the kernel density estimation or χ 2 Distribution determination control limits T 2 and Q.

[0042] The static principal component d r (t) and the static residual e r (t) is obtained by the following steps:

[0043] (1) For a new sample x t Extract dynamic latent variables d(t):

[0044] d(t)=R T x t

[0045] R=W(P T W) -1

[0046] Among them, R is used to extract the new sample x t Extract the orthogonal projection matrix of dynamic latent variables, W is the feature extraction matrix, and P is the load matrix corresponding to multiple dynamic latent variables;

[0047] (2) Based on the extracted dynamic latent variable d(t), an autoregressive model is constructed for prediction:

[0048]

[0049] Among them, v(t) is the prediction residual, and the autoregressive parameter Θ i The static component e(t) is estimated by least squares:

[0050]

[0051] (3) Establish a PCA model for the static component e(t) to obtain the static principal component and static residual:

[0052] d r (t) = P r T e(t)

[0053] e r (t)=(IP r P r T )e(t)

[0054] Among them, P r is the principal component projection matrix, and I is the identity matrix.

[0055] The step 5) comprises the following steps:

[0056] 5.1) For new sample x t Perform normalization to obtain x′ t And calculate its membership degree U for each sub-condition iq , q is the center point of each cluster, and the sub-model j corresponding to the highest membership degree is selected;

[0057] 5.2) The processed new sample x′ t Bring in the dynamic internal principal component analysis model of working condition j and calculate T t 2 and Q t Statistics, if T t 2 >T 2 or Q t >Q will trigger an alarm.

[0058] The present invention has the following beneficial effects and advantages:

[0059] In actual production, cartridge production processes often experience varying operating modes due to factors such as raw material batch differences, equipment status fluctuations, and ambient temperature variations. Using DDFCM technology to locally group data allows data from similar operating conditions to be grouped together, making each local model more closely aligned with the current actual production state. This effectively addresses operating condition variations caused by differences in the production environment and equipment.

[0060] 2. During the production process, key process parameters often exhibit significant temporal fluctuations and hysteresis effects. For example, a small fluctuation in the coating tank temperature can have a cumulative impact on subsequent roll formation. DiPCA dynamically models each local dataset and extracts dynamic features from the time series. This allows rapid detection of potential anomalies, even when equipment operating conditions are unstable or initial anomalies occur, providing timely warnings to on-site personnel.

[0061] 3. By integrating DDFCM and DiPCA technologies and making full use of the global similarity and local dynamic characteristics of the data, the anomaly monitoring system can maintain high robustness and stable monitoring effects in different production batches and complex production environments, thereby significantly improving the generalization performance of the model.

[0062] 4. This fusion method not only reduces the data dimension but also retains key dynamic features, providing more accurate monitoring statistics and control limits for anomaly monitoring, further enhancing the effect of anomaly monitoring and providing strong guarantees for the stability of the production process and product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0064] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0065] like Figure 1 As shown, a method for monitoring and early warning abnormal working conditions in the precision molding process of a cartridge includes the following steps:

[0066] 1) Collect data from cartridge production process to form sample database D0 = [x1, x2, ... x n ], where n is the number of samples;

[0067] The process data collected in step 1) include process variables such as cartridge coating tank temperature, coating speed, paper feed speed, paper feed force, and coiling speed, as well as inspection indicators related to cartridge quality; the raw data D0 is zero-meaned, standardized, or otherwise preprocessed to obtain a training data set D′ train , used for subsequent clustering and modeling.

[0068] The initial data set D0=[x1, x2,…, x n ],n=N,X∈R N×M , where X is the sample feature, N is the number of samples, and M is the sample dimension. The number of clusters M is determined based on the actual working conditions and clustering effectiveness indicators.

[0069] 2) Normalize the historical production data in the sample database D0 to obtain the training data set D′ train , the data normalization formula is as follows:

[0070]

[0071] where μ i is its mean, σ i The processed data constitutes the training data set, which lays the foundation for subsequent clustering and dynamic modeling;

[0072] 3) Using the diffusion distance fuzzy clustering (DDFCM) algorithm, the global similarity of the data is measured by diffusion distance, and the training set is divided into M cluster sub-datasets D′ train =[D′ train1 ,D′ train2 ,...,D′ trainM ]. The diffusion distance is defined as:

[0073]

[0074] Where i, j, l = 1, 2...n, φ0(x) is the penalty difference factor corresponding to different sample point density areas, p2(x i ,x j ) is point x i To point x j The 2-step transition probability, deg(x l ) is the sample local density weight;

[0075] The fuzzy clustering DDFCM algorithm based on diffusion distance in step 3) obtains the similarity of the data in the global structure by constructing a diffusion distance measurement matrix between samples, and classifies samples with high similarity into the same fuzzy class. The specific features of the fuzzy clustering DDFCM algorithm based on diffusion distance are as follows:

[0076]

[0077] Where J m is the model objective function, m is the membership degree U pq The weight factor, is the 2-step transfer probability from point p to point q, i.e., the diffusion distance. When the objective function is minimized, the membership function U pq and cluster center c q The expression is as follows:

[0078]

[0079] 4) For each sub-dataset D′ traini Construct a dynamic internal principal component analysis (DiPCA) model. For the time series vector x t , generate dynamic latent variables d(t)=w by projecting vector w T x t , whose predicted values ​​satisfy the multi-order autoregressive constraints: where β=[β1,...,β s ] T is the weight coefficient vector, s is the preset lag order;

[0080] 5) Extracting time series features by optimizing the objective function of the local dynamic internal principal component analysis (DiPCA) model:

[0081]

[0082] st||w||=1,||β||=1

[0083] in represents the Kronecker product where

[0084] Z s=[X1,X2,…,X s ]

[0085] X i =[x i ,x i+1 ,…,x(N+i-1)] T

[0086] 6) After extracting the dynamic latent variable feature d = Xw, iteratively update X as follows:

[0087] X:=X-dp T

[0088] Where p = X T d / (d T d) is the load vector. Through the above method, multiple dynamic latent variables can be extracted, and the corresponding load matrix P and feature extraction matrix W can be obtained.

[0089] 7) After extracting the dynamic latent variables, calculate T 2 (t) and Q(t) statistics and their control limits:

[0090] d r (t) = P r T e(t)

[0091] e r (t)=(IP r P r T )e(t)

[0092]

[0093] Among them, P r is the principal component projection matrix, I is the identity matrix, d r (t) is the static principal component, e r (t) is the static residual. T 2 The (t) statistic measures the degree of deviation of the sample in the principal component subspace (score space), reflecting the "internal" changes of the system as a whole from the normal state. The SPE (Q(t) statistic) measures the residual size of the sample outside the reconstruction space and is sensitive to "external" anomalies that cannot be captured by the principal component subspace. Control limit T 2 and Q by kernel density estimation or χ 2 Distribution determination;

[0094] In step 7), abnormality monitoring is achieved through dynamic-static combined indicators:

[0095] 1) For a new sample x t Extract the dynamic latent variable d(t) as follows: d(t) = R T xt , where R=W(P T W) -1 .

[0096] 2) Based on the extracted dynamic latent variable d(t), an autoregressive model is constructed for prediction, which is expressed as:

[0097]

[0098] Where v(t) is the prediction residual, and the autoregressive parameter Θ i It can be estimated by least squares. The static component e(t) can be expressed as

[0099] 3) By building a PCA model for the static component e(t), we can obtain the static principal component and static residual, which can be expressed as

[0100] d r (t) = P r T e(t)

[0101] e r (t)=(IP r P r T )e(t)

[0102] Among them, P r is the principal component projection matrix, and I is the identity matrix.

[0103] 8) For the new sample x t Perform normalization to obtain x′ t And calculate its membership degree U for each sub-condition iq , q is the center of each cluster. Select the sub-model j with the highest membership;

[0104] 9) The processed new sample x′ t Substitute the DiPCA model of working condition j and calculate T t 2 and Q t Statistics, if T t 2 >T 2 or Q t >Q will trigger an alarm.

[0105] It can be seen from the above technical solution that compared with the prior art, the present invention discloses a method for abnormality monitoring of the cartridge coating, rolling and cutting processes. The method first performs global similarity measurement and local clustering on the normalized multidimensional process data through the diffusion distance-based fuzzy clustering (DDFCM) technology, thereby automatically dividing the key sample groups of different working conditions; then, for each cluster sub-data set, the dynamic internal principal component analysis (DiPCA) method is used to extract dynamic latent variables under multi-order autoregressive constraints, effectively capturing the temporal variation characteristics of the coating, rolling and cutting processes, and calculating the T2 and SPE statistics based on the extracted dynamic and static principal components, and determining the corresponding control limits through kernel density estimation or statistical distribution, thereby realizing online abnormality discrimination and real-time alarm for new samples.

[0106] The various steps of the method of the present invention are closely linked to each other and cannot be implemented independently without any link, and the relational terms such as "first" and "second" are only used to distinguish different entities or operations, and do not imply the order of their execution; the terms "include", "comprise" and any variations thereof are intended to cover non-exclusive inclusion, that is, the process, method or system may include other elements or inherent characteristics in addition to the explicit elements, further enhancing the adaptability and robustness of the present invention in a changing production environment.

[0107] The above description is merely an example of a feasible embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modification, equivalent replacement, improvement or expansion made within the spirit and principle of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A method for monitoring and early warning abnormal working conditions during the precision molding process of a cartridge, characterized in that: The following steps are involved: 1) Collect data from the cartridge production process and normalize it to obtain the training data set D′ train ; 2) Use the diffusion distance-based fuzzy clustering method to divide the training data set into stages and obtain M cluster sub-datasets D t ' rain =[D t ' rain1 ,D t ' rain2 ,...,D t ' rainM ]; 3) A dynamic internal principal component analysis model is constructed for each cluster sub-dataset, and its objective function is optimized to extract dynamic latent variables; 4) Based on dynamic latent variables, calculate the core statistics T of process monitoring i 2 and Q i , and obtain the control limit T 2 and Q; 5) Repeat steps 3) and 4) until the dynamic internal principal component analysis model is constructed for all M cluster sub-datasets and the corresponding control limits are calculated; 6) Calculate the new sample x t The membership degree of the new sample x is analyzed using the dynamic internal principal component analysis model corresponding to the membership degree. t Conduct abnormal monitoring.

2. The method for monitoring and warning abnormal working conditions during the precision molding process of a cartridge according to claim 1 is characterized in that: The step 2) comprises the following steps: 2.1) Calculate the diffusion distance D(x) between samples i ,x j ): Among them, variables i, j, l = 1, 2...n, φ0(x) is the penalty difference factor corresponding to different areas of sample density, p2(x i ,x j ) is point x i To point x j The 2-step transition probability, deg(x l ) is the sample local density weight; 2.2) Obtaining the degree of data membership in the global structure based on the diffusion distance fuzzy clustering method: Among them, J m is the objective function of the clustering algorithm, and m is the membership degree U pq The weight factor, is the 2-step transfer probability from point p to point q, i.e., the diffusion distance. When the objective function is minimized, the membership function U pq and cluster center c q for: 2.3) Classify samples with membership within the same threshold range into the same fuzzy class.

3. The method for monitoring and warning abnormal working conditions during the precision molding process of a cartridge according to claim 1 is characterized in that: The step 3) comprises the following steps: 3.1) Define dynamic latent variable extraction rules: For a time series vector x t , generate dynamic latent variables d(t)=w by projecting vector w T x t , whose predicted values ​​satisfy the multi-order autoregressive constraints: i=1,2,...,s, where β=[β1,...,β s ] T is the weight coefficient vector, s is the preset lag order; 3.2) Construct dynamic optimization objective function: st||w||=1,||β||=1 in, represents the Kronecker product, where: Z s =[X1,X2,...,X s ] X i =[x i ,x i+1 ,…,x(N+i-1)] T Among them, X i represents the data submatrix intercepted at the i-th moment in the time series, Z s Indicates multiple Xs i The collection after horizontal merger; 3.3) After extracting the dynamic latent variable feature d = Xw, iteratively update X: X:=X-dp T Where p = X T d / (d T d) is the load vector; 3.4) After step 3.3), multiple dynamic latent variables are extracted, and the corresponding load matrix P and feature extraction matrix W are obtained.

4. The method for monitoring and warning abnormal working conditions during the precision molding process of a cartridge according to claim 1 is characterized in that: The step 4) is specifically as follows: d r (t)=P r T e(t) e r (t)=(I-P r P r T )e(t) Q(t)=‖e r (t)‖ 2 =e(t) T (I-P r P r T )e(t) Among them, P r is the principal component projection matrix, I is the identity matrix, d r (t) is the static principal component, e r (t) is the static residual, e(t) is the static component, T 2 (t) is a statistic that measures the degree of deviation of the sample in the principal component subspace, I is the unit matrix, Q(t) represents the Q statistic calculated at time t, and finally the kernel density estimation or χ 2 Distribution determination control limits T 2 and Q.

5. The method for monitoring and early warning abnormal working conditions during the precision molding process of a cartridge according to claim 4 is characterized in that: The static principal component d r (t) and the static residual e r (t) is obtained by the following steps: (1) For a new sample x t Extract dynamic latent variables d(t): d(t)=R T x t R=W(P T W) -1 Among them, R is used to extract the new sample x t Extract the orthogonal projection matrix of dynamic latent variables, W is the feature extraction matrix, and P is the load matrix corresponding to multiple dynamic latent variables; (2) Based on the extracted dynamic latent variable d(t), an autoregressive model is constructed for prediction: Among them, v(t) is the prediction residual, and the autoregressive parameter Θ i The static component e(t) is estimated by least squares: (3) Establish a PCA model for the static component e(t) to obtain the static principal component and static residual: d r (t)=P r T e(t) e r (t)=(I-P r P r T )e(t) Among them, P r is the principal component projection matrix, and I is the identity matrix.

6. The method for monitoring and warning abnormal working conditions during the precision molding process of a cartridge according to claim 1 is characterized in that: The step 5) comprises the following steps: 5.1) For new sample x t Perform standardization to obtain x t 'And calculate its membership degree U for each sub-condition iq , q is the center point of each cluster, and the sub-model j corresponding to the highest membership degree is selected; 5.2) The processed new sample x t 'Introduce the dynamic internal principal component analysis model of working condition j and calculate T t 2 and Q t Statistics, if T t 2 >T 2 or Q t >Q will trigger an alarm.

Images