An improved particle filtering and convolution network-based engraving machine anomaly detection method

Abstract

The application discloses a kind of based on improved particle filtering and convolution network's engraving machine anomaly detection method, and is divided into two stages of offline training and anomaly detection.In the offline training stage, first, collect the data of engraving machine operation and determine the system model structure under normal working condition, then introduce improved particle filtering algorithm to obtain system state variable estimation value, and using expectation maximization algorithm iteratively updates model parameters, obtains the accurate mathematical model of engraving machine system, finally, the residual information of model is used to train one-dimensional residual convolution neural network, realize the two classification of system operation data.In the anomaly detection stage, the noise estimation value sequence is obtained by particle filtering using the mathematical model to the operation data, and finally whether the running data is abnormal is judged by the classifier trained offline.The application can obtain accurate mathematical model of system and mine the correlation of data in time dimension, improve the precision of engraving machine operation data anomaly detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of engraving machine operation anomaly detection, and particularly relates to an engraving machine operation data anomaly detection method based on improved particle filtering and convolutional network. BACKGROUND

[0002] At present, engraving machines are widely used in mold, tobacco, automobile, model making, mechanical processing and other industries. With the rapid development of manufacturing industry and network technology, higher requirements are put forward for the safety performance of engraving machines. On the one hand, the engraving machine system still faces traditional functional safety threats such as equipment failure. On the other hand, although the application of network technology brings advantages such as resource sharing and convenient management, it breaks the closed nature of the engraving machine system, making the system face information security threats such as viruses and hacker attacks. If the equipment cannot be detected and maintained in time after an anomaly occurs, it may cause the overall performance of the system to change, the service life of the equipment to be shortened, and other consequences. Anomaly detection technology can detect anomalies at an early stage to reduce or avoid accidents, which is of great significance to the safety and reliability of the operation of the equipment.

[0003] A traditional anomaly detection method is to establish a model of the system in a normal state, and to detect anomalies based on whether it deviates significantly from the normal model. Establishing a system model to accurately describe the behavior characteristics of the system in a normal state, and establishing a classification model to accurately distinguish between normal data and abnormal data are two important links of this method. In the stage of establishing the system model, the method of combining particle filtering with expectation maximization algorithm can be used to evaluate the model parameters. This method can reduce the influence of observation noise on state estimation, obtain accurate state estimation value, and then obtain more appropriate parameters in the evaluation of model parameters to improve the accuracy of the system model. However, the standard particle filtering algorithm does not use the observation information of future time when estimating the variable at a certain time, resulting in insufficient filtering accuracy. By introducing the forward-backward algorithm, the observation information of future time can be fully utilized to improve the filtering accuracy, but the calculation time will increase quadratically with the increase of the number of particles, and there is a defect that a large number of particles cannot be set.

[0004] In the classifier training stage, the traditional threshold-based method only considers the residual information at a single time point, and cannot fully exploit the correlation of residual data in the time dimension, making it difficult to handle context anomalies in specific working conditions. Generally, the value of a point in the time series data of an engraving task is within the normal range from the isolated perspective, and cannot constitute an anomaly alone. When the values of other points around the point are given as the context environment, the point deviates from the normal condition under this environment, thus constituting a context anomaly. The context environment can reflect key information such as equipment operation conditions, and the traditional threshold-based method cannot utilize these context information to improve the accuracy of operation data anomaly detection.

[0005] In view of the shortcomings of the above method, the present application proposes an engraving machine anomaly detection method based on improved particle filtering and convolutional neural network. The improved particle filtering algorithm is combined with the expectation maximization algorithm to evaluate the model parameters, so that the algorithm can fuse more observation information to improve the filtering accuracy, while overcoming the defect that the calculation time increases quadratically with the number of particles. In addition, by training the residual neural network to fully exploit the correlation of data in the time dimension, the implicit correlation features can be extracted and used for classifier training, and a binary classification model that can accurately distinguish normal data and abnormal data is obtained, improving the accuracy of anomaly detection. SUMMARY

[0006] In order to overcome the shortcomings of the prior art, the present application proposes an engraving machine anomaly detection method based on improved particle filtering and convolutional network. In the evaluation of model parameters, the improved particle filtering algorithm is used to accurately estimate the system state variables, which not only improves the filtering accuracy, but also overcomes the defect that the calculation time increases quadratically with the number of particles; then the expectation maximization algorithm is used to iteratively update the model parameters to obtain an accurate mathematical model of the engraving machine system. In the decision-making stage, a one-dimensional residual convolutional neural network is used to fully exploit the correlation of data in the time dimension, achieving more accurate and reliable anomaly detection.

[0007] The technical solution adopted by the present application to solve its technical problems is:

[0008] An engraving machine operation data anomaly detection method based on improved particle filtering and convolutional neural network, in the offline training stage, the improved particle filtering algorithm and the expectation maximization algorithm are combined to iteratively update the parameters of the mathematical model of the engraving machine system, and the noise estimation value sequence after particle filtering is used as input, and a one-dimensional residual convolution is used to train an anomaly detection classifier,

[0009] The method steps are as follows:

[0010] 1) Collect the operation data of the engraving machine system under normal working conditions, wherein the time series length L, the input data Observation data

[0011] 2) Determine the mathematical model structure of the engraving machine system under normal working conditions and initialize the parameters;

[0012] 3) For data under normal working conditions, use the improved particle filtering algorithm to obtain the estimation value sequence of the state variable at each time

[0013] 4) Use the state variable estimation value sequence obtained by the algorithm in step 3) to train the mathematical model of the engraving machine system under normal working conditions;

[0014] 5) Collect the data of the engraving machine system under normal and abnormal conditions, and perform offline training of the classifier to obtain a binary classification model capable of distinguishing between normal and abnormal data;

[0015] 6) Perform particle filtering on the running data of the to-be-detected engraving machine using the mathematical model of the system obtained in step 4) to calculate the estimated value of the noise, and then use the classification model obtained in step 5) to distinguish between normal and abnormal data.

[0016] Further, the mathematical model structure of the engraving machine system under normal conditions in step 2) and the initialization of the parameters are as follows:

[0017] 2.1) Determine the single-step prediction model of the state variable of the engraving machine,

[0018] x t+1 =f(x t ,u t )+q t ,t=1,2,…,L (1)

[0019] where f(x t ,u t ) represents the single-step prediction function of the state variable of the engraving machine under noise-free conditions, which is determined according to the specific situation, such as f(x t ,u t )=Ax t +Bu t , where A and B are coefficient matrices, and the numerical values are to be identified, x t represents the state variable of the engraving machine at time t, q t is a system noise following a Gaussian distribution with a mean of 0 and a covariance matrix Q, and the elements of Q on the non-main diagonal are 0;

[0020] 2.2) Determine the observation model of the state variable of the engraving machine,

[0021] y t =h(x t ,u t )+r t ,t=1,2,…,L (2)

[0022] where h(x t ,u t ) represents the state variable observation function of the engraving machine under noise-free conditions, which is determined according to the specific situation, such as h(x t ,u t )=Cx t +Du t , where C and D are coefficient matrices, and the numerical values are to be identified, y tr t is the observation noise obeying Gaussian distribution with mean 0 and covariance matrix R, and the non-diagonal elements of R are 0;

[0023] 2.3) Set Q, R and the initial value of x1 according to prior knowledge.

[0024] Further, the improved particle filter algorithm uses a method whose calculation time increases linearly with the number of particles to fuse the observation information at future time, and uses a circular array to store the ancestor particle information of the particles to improve the calculation efficiency,

[0025] Step 3) The improved particle filter algorithm is used to obtain the estimated value sequence of the state variable at each time The specific implementation steps are as follows:

[0026] 3.1) Set the number of particles as N, and the value of the i-th particle at time t as The weight is At time t = 1, the initial value of all particles is The weight is

[0027] 3.2) For t = 2, 3, …, L, the following steps are sequentially executed:

[0028] 3.2.1) Sample the value of each particle at time t

[0029]

[0030] denotes the probability distribution of the state variable at time t under the condition that is the value of the i-th particle after resampling at time t-1, and let

[0031] 3.2.2) Update the weight according to the observation value, and set the observation value at time t as y t Since the initial weight of the particle and the weight after resampling are both 1 / N, the weight of the particle at the previous time can not be considered during calculation, and the weight of the particle with subscript i at time t is obtained

[0032]

[0033] where is the probability that the particle value is t under the condition that y is observed;

[0034] ​3.2.3) save the weight of the particle at the current time and the value of the particle at the current time;

[0035] 3.2.4) generate a new particle population using the standard resampling method; let the ith old particle before resampling be the jth new particle after resampling be If is copied from , then is called an ancestor particle of , and all ancestor particles of are also ancestor particles of ; in addition, if a particle at time t1is an ancestor particle of , it is called an ancestor of at time t1;

[0036] 3.2.5) for each particle after resampling, use a cyclic array of length l step to save the superscript information of its ancestor at times t, t-1,..., max(1, t-l step +1), each time after resampling, the new particle copies the cyclic array of the old particle, and then assigns the (t-2) % l step value in the array to the superscript i of the old particle p t (i) , and the cyclic array of the new particle is obtained;

[0037] 3.3) backward update the particle weight, let the step size of the backward update be l step , for t = l step +1, l step +2,..., L, perform the following operations:

[0038] According to the weight of the particle at time t , backward update the weight of its ancestor at time t-l step , let the weight obtained after the backward update be

[0039]

[0040] F j (t) represents the superscript of the ancestor particle of the particle with superscript j at time t at time t-l step ; if a certain particle is not an ancestor of any particle after l step time, then its weight is 0;

[0041] 3.4) for t = 1, 2,..., l step , let

[0042] 3.5) Obtain the estimated value of the state variable according to the particle value and weight at each time point,

[0043]

[0044] wherein is the estimated value of the state variable at time t;

[0045] Further, the improved particle filtering algorithm is combined with the expectation maximization algorithm to obtain a method for training model parameters by combining the improved particle filtering algorithm with the expectation maximization algorithm. After pre-training of the model, the estimated value of the system state variable is obtained by using the improved particle filtering algorithm, and the model parameters are continuously iteratively updated by using the expectation maximization algorithm.

[0046] Step 4) The mathematical model of the training engraving machine in normal working condition is implemented as follows:

[0047] 4.1) Pre-train the single-step prediction model parameters. For the estimated value sequence of the state variable at each time point obtained in step 3) Set the loss function as formula (7),

[0048]

[0049] Use the gradient descent method to iteratively update the single-step prediction model parameters of the state variable and use them as the initial values of step 4.2);

[0050] 4.2) Update the model parameters by using the method of combining the improved particle filtering algorithm with the expectation maximization algorithm. Update the single-step prediction model, the state variable observation model, the process noise variance Q, and the observation noise variance R by using the gradient descent method. Before each calculation of the gradient, perform particle filtering once according to step 3) to update the estimated value sequence of the state variable The loss function is formula (8),

[0051]

[0052] wherein D x is the dimension of the variable x, wherein D y is the dimension of the variable y, is the data of the dth dimension of the state variable estimated value at time t+1, is the data of the dth dimension of , Q d is the variance of the dth dimension of the process noise, y t,d is the data of the jth dimension of the observation value at time t, is the data of the jth dimension of , R d is the variance of the dth dimension of the process noise.

[0053] Further, the step 5) described above is offline training to obtain a binary classification model for distinguishing normal data and abnormal data, and the specific implementation steps are as follows:

[0054] 5.1) The data collected in step 1) is divided into training set and test set, each sample is a continuous time of sensor time series data, and whether there is abnormal data in the time is used as a classification label;

[0055] 5.2) Use the improved particle filter algorithm described in step 3) to process the training set data to obtain the estimated value of the state variable, and obtain the process noise estimation sequence and the observation noise estimation sequence

[0056] 5.3) Use the noise sequence obtained in step 5.2) as input to train a binary classification model using a one-dimensional residual convolution network.

[0057] Further, the step 6) described above is to obtain a binary classification model for distinguishing normal data and abnormal data, and the specific implementation steps are as follows:

[0058] 6.1) Collect the input data and observation data of the engraving machine;

[0059] 6.2) Use the model obtained in step 4) to perform the improved particle filter algorithm described in step 3) on the collected data to obtain a sequence of noise estimation values;

[0060] 6.3) Use the residual convolution network obtained in step 5) to determine whether the sequence obtained in step 6.2) is in an abnormal state.

[0061] The method has the following advantages:

[0062] 1) Compared with the standard particle filter method, the information of future time can be fused when estimating the state at a certain time, and a more accurate filtering result is obtained;

[0063] 2) The time complexity of the calculation is reduced, and the situation that the calculation time increases quadratically with the number of particles is avoided, so that a larger number of particles can be set for particle filtering;

[0064] 3) A one-dimensional residual convolution network is used to judge the abnormality, which can mine the correlation of data in the time dimension, and is not limited to single-point information, and the accuracy of the classifier is improved. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 The flowchart of the steps in the model training stage and the abnormality detection stage;

[0066] Figure 2To improve the particle filtering step flow chart;

[0067] Figure 3 For one-dimensional residual convolutional neural network structure diagram. Specific embodiments

[0068] The present application will be further described in detail below with reference to examples and accompanying drawings.

[0069] As Figure 1 shown in a kind of engraving machine abnormality detection method based on improved particle filtering and convolution network, comprising the following steps:

[0070] 1) collection engraving machine normal operating condition under the running data, wherein time series length L, 2-dimensional input data sequence 2-dimensional observation data sequence Wherein, normal operating condition refers to the expected operating condition of equipment under specified execution task;For deviating from the expected operating condition, the operating condition of the appearance fault or the premonition before fault, all be considered to belong to non-normal operating condition, i.e. abnormal operating condition;

[0071] 2) determine the mathematical model structure under normal operating condition of system and initialize parameter, including the following steps:

[0072] 2.1) determine the single-step prediction model of engraving machine state variable,

[0073] x t+1 =Ax t +Bu t +q t ,t=1,2,…,L (1)

[0074] Wherein, A 44 4 rows 4 columns coefficient matrix, B 42 4 rows 2 columns coefficient matrix, the numerical value of parameter matrix is to be identified, x t Indicate the 4-dimensional state variable of engraving machine at t time, q t It is the system noise that obeys Gaussian distribution with mean 0 and covariance matrix Q, and the elements of Q are 0 on the main diagonal;

[0075] 2.2) determine the observation model of engraving machine state variable,

[0076] y t =Cx t +Du t +r t ,t=1,2,…,L (2)

[0077] Wherein, C 24 2 rows 4 columns coefficient matrix, D 22 2 rows 2 columns coefficient matrix, the numerical value of parameter matrix is to be identified, yt y(t) represents the observation value of the engraving machine at time t, r t is the observation noise obeying Gaussian distribution with mean 0 and covariance matrix R, and the non-diagonal elements of R are 0;

[0078] 2.3) Set Q, R and the initial value of x1 according to prior knowledge;

[0079] 3) For the data under normal working condition, use the improved particle filter algorithm to obtain the estimated value sequence of state variable at each time The specific implementation steps are as follows:

[0080] 3.1) Set the number of particles N = 2000, and the value of the i-th particle at time t is The weight is At time t = 1, the initial value of all particles is The weight is

[0081] 3.2) For t = 2, 3,..., L, execute the following steps in turn:

[0082] 3.2.1) Sample the value of each particle at time t

[0083]

[0084] denotes the probability distribution of the state variable at time t under the condition of is the value of the i-th particle after resampling at time t-1, and let

[0085] 3.2.2) Update the weight according to the observation value, and set the observation value at time t as y t Since the initial weight of the particle and the weight after resampling are both 1 / N, the weight of the i-th particle at time t can be obtained without considering the weight of the particle at the previous time

[0086]

[0087] where is the probability that the particle value is t

[0088] 3.2.3) Save the particle weight at the current time and the particle value at the current time;

[0089] 3.2.4) Generate a new particle population using the standard resampling method; set the i-th old particle before resampling as ​​​The j-th new particle after resampling is like It is by If it is copied, then it is called yes The ancestral particle, and p t (i) All ancestral particles are also The ancestral particle; furthermore, if a particle at time t1 is The ancestral particle, called Ancestor at time t1;

[0090] 3.2.5) For each particle after resampling, use a length of l step The circular array stores its values ​​in t, t-1, ..., max(1, tl) step +1) The superscript information of the ancestor at time step, after each resampling, the new particle copies the circular array of its old particles, and then the (t-2)%l of the array is used. step The value is assigned to the old particles before resampling. By using the superscript i, we can obtain the circular array of the new particles;

[0091] 3.3) Backward update of particle weights, assuming the step size of the backward update is l. step For t=l step +1,l step +2,…,L, perform the following operations:

[0092] Based on the particle weights at time t Backward update its tl step The weights of the ancestors at time step , let the weights obtained after the backward update be...

[0093]

[0094] F j (t) represents the particle with superscript j at time tl. step The superscript of the ancestor particle at time 1, if a particle is not l step If a particle is the ancestor of any particle after a certain time, then its weight is 0.

[0095] 3.4) For t = 1, 2, ..., l step ,make

[0096] 3.5) Based on the particle value and weight at each time step, obtain the estimated values ​​of the state variables.

[0097]

[0098] in is the estimated value of the state variable at time t;

[0099] 4) Using the state variable estimation value sequence obtained by the method of step 3), training the mathematical model of the engraving machine system under normal working condition, the specific steps are as follows:

[0100] 4.1) Pre-training the single-step prediction model parameters, for the state variable estimation value sequence obtained in step 3) Set the loss function as formula (7),

[0101]

[0102] Use the gradient descent method to iteratively update the parameters of the coefficient matrix A, B and use them as the initial values of step 4.2);

[0103] 4.2) Update the model parameters using the improved particle filtering algorithm combined with the expectation maximization algorithm, and use the gradient descent method to update the coefficient matrix A, B, C, D, process noise variance Q and observation noise variance R. Before each gradient calculation, perform a particle filtering according to step 3) once to update the estimated value of the state variable The loss function is as formula (8),

[0104]

[0105] Where is the d-dimensional data of the state variable estimation value at time t+1, is the d-dimensional data of , Q d is the variance of the d-dimensional process noise, y t,d is the d-dimensional data of the observation value at time i, is the d-dimensional data of , R d is the variance of the d-dimensional process noise;

[0106] 5) Collect data of the engraving machine system under normal and abnormal working conditions, and perform offline training of the classifier to obtain a binary classification model that can distinguish normal data and abnormal data, the specific implementation steps are as follows:

[0107] 5.1) The data collected in step 1) is divided into training set and test set when the engraving machine is running in normal and abnormal state, each sample is a time series data of sensor in a short time, and whether there is abnormal data in a short time is used as the classification label;

[0108] 5.2) Use the improved particle filtering algorithm in step 3) to process the training set data to obtain the estimated value of the state variable, and obtain the process noise estimation value sequence according to formula (1) and formula (2) and the observation noise estimation sequence

[0109] 5.3) Using the noise sequence obtained in step 5.2) as input, train a binary classification model using a one-dimensional residual convolutional network;

[0110] The residual convolutional network structure is shown in FIG. 5, the input first enters a one-dimensional convolutional layer, and then sequentially enters 6 BasicBlock blocks. After passing through 2 BasicBlock blocks, the channel number is doubled when entering the first convolutional layer of the next BasicBlock block, and the sequence length is reduced to half by using skip-convolution. Finally, the maximum value of the time sequence of each channel is taken, and a one-dimensional linear layer is entered to obtain the probability of being in an abnormal working condition. Figure 3

[0111] 6) Using the mathematical model of the system obtained in step 4) to perform particle filtering on the running data of the to-be-detected engraving machine, calculate the estimated value of the noise, and then use the classification model obtained in step 5) to distinguish normal data and abnormal data, the specific implementation steps are as follows:

[0112] 6.1) Collect the input data and observation data of the engraving machine;

[0113] 6.2) Use the model obtained in step 4) to perform the improved particle filtering algorithm described in step 3) on the collected data to obtain a sequence of noise estimate values;

[0114] 6.3) Use the residual convolutional network obtained in step 5) to determine whether the sequence obtained in step 6.2) is in an abnormal state.

[0115] The above specific embodiments only describe the preferred embodiments of the present application, and do not limit the protection scope of the present application. Any modifications, substitutions and improvements of the technical solutions of the present application made by those skilled in the art according to the description and drawings of the present application without departing from the design concept and spirit of the present application shall belong to the protection scope of the present application.​

Claims

1. An improved particle filter and convolutional network based engraver anomaly detection method, characterized in that, The method comprises the following steps: 1) Collecting the running data of the engraving machine system under normal working conditions, wherein the time series length L, the input data Observation data Wherein, the normal working condition refers to the expected running condition of the equipment under the specified execution task; the working condition deviating from the expected running condition, the fault or the premonition before the fault is considered as the non-normal working condition, i.e. the abnormal working condition; 2) determining the mathematical model structure of the engraving machine system under normal working conditions and initializing parameters; 3) For the data under normal operating conditions, the improved particle filter algorithm is used to obtain the estimated value sequence of the state variable at each time Step 3) The specific steps are as follows: 3.1) Let the number of particles be N, and the value of the ith particle at time t be The weights are At time t = 1, the initial values of all particles are The weights 3.2) For t = 2, 3, …, L, the following steps are executed in turn: 3.2.1) Sampling at time t to get the value of each particle denotes the probability distribution of the state variable at time t under the condition that is the value of the i-th particle after resampling at time t-1, and let ​ 3.2.2) Update the weight according to the observation value, set the observation value at time t as y t Since the initial weight of the particle and the weight after resampling are both 1 / N, the weight of the particle marked as i at time t can be obtained without considering the weight of the particle at the previous time where is the probability that the particle value is t given that y t was observed. 3.2.3) Save the particle weight of the current time and the particle value of the current time; 3.2.4) Generate a new population of particles using a standard resampling method; let the ith old particle before resampling be the jth new particle after resampling be If is copied from then we say that is an ancestor particle of and all ancestor particles of are also ancestor particles of ; furthermore, if a particle at time t1is an ancestor particle of we say that it is an ancestor at time t1of . 3.2.5) For each particle after resampling, a circular array of length l step is used to keep the superscript information of its ancestors at times t, t-1,..., max(1, t-l step +1). After each resampling, the new particle copies the circular array of its old particle, and then assigns the (t-2)th value in the array to the superscript i of the old particle before resampling, and thus obtains the circular array of the new particle. step The superscript i of the old particle before resampling, and thus obtains the circular array of the new particle.​ 3.3) Backward update the particle weights to incorporate the observation information at future time, let the step size of backward update be l step For t = 1 step + 1, 1 step + 2,..., L, do the following: According to the weight of the particle at time t The weight of the ancestor at time t-l is updated backward step The weight of the ancestor at time t-l is updated backward F j (t) denotes the superscript of the ancestor particle of the particle with superscript j at time t - 1 step if a particle is not an ancestor of any particle after time 1 step its weight is 0; 3.4) for t = 1,2,..., l step , let 3.5) According to the particle value and weight of each time, the estimated value of the state variable is obtained, wherein is the estimate of the state variable at time t; 4) Using the state variable estimation value sequence obtained by the algorithm in step 3), combined with the expectation maximization algorithm, the mathematical model of the engraving machine system under normal working conditions is iteratively trained; 5) Collect data of the engraving machine system under normal working conditions and abnormal working conditions, and perform offline training of the classifier to obtain a binary classification model capable of distinguishing normal data and abnormal data; 6) Use the mathematical model of the engraving machine system obtained in step 4) to perform particle filtering on the running data of the engraving machine system to be detected, calculate the estimated value of the noise, and then use the classification model obtained in step 5) to distinguish normal data and abnormal data.

2. The method according to claim 1, wherein, The specific steps of step 2) are as follows: 2.1) Determine the single-step prediction model of the state variable of the engraving machine, x t+1 = f(x t , u t ) + q t , t = 1, 2,..., L where f(x t ,u t ) represents a single-step prediction function of the state variable of the engraving machine in the case of no noise, the form of which is determined according to specific conditions, such as f(x t ,u t ) = Ax t + Bu t , where A and B are coefficient matrices, and the numerical values are to be identified, x t represents the state variable of the engraving machine at time t, and q t is system noise that obeys a Gaussian distribution with a mean of 0 and a covariance matrix Q, and the elements of Q that are not on the main diagonal are 0; 2.2) Determine the observation model of the state variable of the engraving machine, y t = h(x t , u t ) + r t , t = 1, 2,..., L where h(x t ,u t ) represents the state variable observation function of the engraving machine in the case of no noise, the form of which is determined according to the specific situation, such as h(x t ,u t ) = Cx t + Du t , wherein C and D are coefficient matrices, and the numerical values are to be identified, y t represents the observation value of the engraving machine at time t, and r t is observation noise subject to a Gaussian distribution with a mean of 0 and a covariance matrix R, and the elements on the non-main diagonal of R are 0; 2.3) Set Q, R and x1 initial value according to prior knowledge.

3. The method of claim 1, wherein, The specific steps of step 4) are as follows: 4.1) Pre-training of the single-step prediction model parameters, for the sequence of estimates of the state variables of the engraver at each time instant The loss function as in equation (5) is set where f(x t ,u t ) represents the one-step prediction function of the state variable of the engraving machine under the condition of no noise, and h(x t ,u t ) represents the observation function of the state variable of the engraving machine under the condition of no noise; the parameters of the one-step prediction model of the state variable are iteratively updated using the gradient descent method and used as the initial value of step 3.2); 4.2) iteratively update the model parameters using the method combining improved particle filter with expectation maximization, in each iteration, use the improved particle filter algorithm described in step 2) to update the sequence of state variable estimates Then update the one-step prediction model, the state variable observation model, the process noise variance Q, the observation noise variance R using gradient descent method, the loss function is as formula (6), where D x is the dimension of the variable x, where D y is the dimension of the variable y, is the data of the dth dimension of the state variable estimate at time t+1, is the data of the dth dimension of , Q d is the variance of the process noise of the dth dimension. y t,d is the data of the jth dimension of the observation at time t, is the data of the jth dimension of , R d is the variance of the process noise of the dth dimension.

4. The method of claim 1, wherein, The specific steps of step 5) are as follows: 5.1) Divide the data collected in step 1) into training set and test set, each sample is a continuous time sensor time series data, and whether there is abnormal data in the time is used as the classification label; 5.2) Using the improved particle filter algorithm of step 3), process the training set data to obtain the estimate of the state variable, and obtain the sequence of process noise estimates from equation (1), equation (2) and the sequence of observation noise estimates 5.3) Use a one-dimensional residual convolution network to train a binary classification model with the noise sequence obtained in step 5.2) as input.

5. The method of claim 1, wherein, The specific steps of step 6) are as follows: 6.1) Collect the input data and observation data of the engraving machine; 6.2) Use the model obtained in step 4) to perform the improved particle filtering algorithm in step 3) on the collected data to obtain a sequence of noise estimation values; 6.3) Use the residual convolution network obtained in step 5) to determine whether the sequence obtained in step 6.2) is in an abnormal state.

Images