Feed production process monitoring method based on ingredient prediction

Abstract

The present invention discloses a feed production process monitoring method based on component prediction, which belongs to the technical field of production process monitoring. The present invention collects mass spectrometry data of feed products in a processing system, calculates the measured content of the component to be tested, and sends a first abnormal signal according to whether the measured content is within a qualified interval. The present invention preprocesses the non-uniform measured component matrix to obtain a uniformly spaced corrected component matrix. Then, a time series state sequence is created in combination with the unloading state of the batching system to predict the predicted content of the component to be tested in the next sampling cycle, and sends a second abnormal signal according to whether the predicted content is within a qualified interval. The present invention can complete the early monitoring of the component to be tested, avoid the waste of raw materials, and improve the monitoring accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of production process monitoring, and in particular to a feed production process monitoring method based on component prediction. Background Art

[0002] Sampling monitoring of the feed production process can ensure the yield rate. The Chinese patent application with publication number CN116993527A discloses a method for optimizing the collection and monitoring of pig feed production data. The method monitors and analyzes the production process of pig feed in real time, finds problems and makes optimization adjustments, thereby improving production efficiency and quality. The feed production line has many process links, and the abnormal values ​​detected at the outlet have a certain lag. After the outlet is abnormal, the raw materials that are already in the batching system and processing system will inevitably be wasted if the ingredients are fed back and adjusted. In order to improve the monitoring timeliness, it is necessary to predict the abnormal values ​​of the next time period in advance. The Chinese patent with publication number CN114756604B discloses a monitoring time series data prediction method based on the Prophet combination model. The method first arranges sensors on site and obtains original time series data, generates complete time series data through the preprocessing layer and regression layer of the combination model, and then substitutes the complete time series data into the Prophet combination model to obtain predicted time series data. This method uses the model regression layer to ensure the availability of sampled data. If the model regression layer does not match the field conditions, the prediction of abnormal values ​​is difficult to achieve the expected goal. Therefore, the prior art needs to be further improved. Summary of the invention

[0003] In order to solve the defects of the above-mentioned prior art, the present invention proposes a feed production process monitoring method based on component prediction. First, the non-uniform sampling data is preprocessed, and then the predicted content of the component to be tested in the next sampling cycle is predicted based on the preprocessed data, so as to complete the early monitoring of the component to be tested, avoid raw material waste and improve monitoring accuracy.

[0004] The technical solution of the present invention is achieved in this way:

[0005] A method for monitoring a feed production process based on component prediction comprises the following steps:

[0006] Step 1: a batching system periodically mixes at least two raw materials, and a processing system receives the mixed raw materials and produces a feed product, wherein the feed product includes a matrix and at least one component to be tested;

[0007] Step 2: Collect mass spectrum data of feed products in the processing system at every sampling period, identify multiple signal peaks according to the mass spectrum data, identify the signal peak of the matrix according to the calibrated content of the matrix ions, and calculate the sampling drift according to the acquisition time of the signal peak;

[0008] Step 3: extracting a signal peak of at least one component to be detected, calculating the mass-to-charge ratio of the component to be detected according to the acquisition time of the signal peak, and calculating the measured content of the component to be detected according to the signal intensity of the signal peak;

[0009] Step 4: If the measured content of at least one component to be tested is outside the qualified range, a first abnormal signal is output and the task is terminated; otherwise, a measured component sequence is generated based on the measured contents of multiple components to be tested;

[0010] Step 5: If the batching system is in the feeding state, the state parameter of the batching system is set to 1, otherwise the state parameter of the batching system is set to 0;

[0011] Step 6: Repeat steps 2 to 5, generate a measured component matrix from the measured component sequences of multiple sampling periods, generate a timing drift sequence from the sampling drifts of multiple sampling periods, and generate a timing state sequence from the state parameters of multiple sampling periods;

[0012] Step 7: Adjust the measured component matrix to a modified component matrix according to the time series drift sequence, and construct a component prediction model according to the modified component matrix and the time series state sequence;

[0013] Step 8: Generate a predicted component sequence for the next sampling period according to the component prediction model, and extract the predicted content of at least one component to be tested in the predicted component sequence;

[0014] Step 9: If the predicted content of at least one component to be tested is outside the qualified range, output a second abnormal signal and end the task, otherwise return to step 2.

[0015] In the present invention, in step 1, the raw materials include at least one or more of corn bran, soybean meal, shell meal, and meat and bone meal, the matrix is ​​one or more of vitamins, proteins or carbohydrates, and the components to be tested are one or more of heavy metals, toxins or pesticides.

[0016] In the present invention, in step 2, the signal intensity of the matrix is ​​determined according to the baseline content of the matrix, a signal peak with the signal intensity is found, the collection time of the signal peak is used as the first collection time, the calibration collection time is calculated according to the mass-to-charge ratio of the matrix ions, and the sampling drift is calculated according to the first collection time and the calibration collection time.

[0017] In the present invention, in step 3, the response ratio of the component to be detected is generated according to the signal intensity of the signal peak of the component to be detected, a response function is fitted, and the response ratio is substituted into the response function to obtain the measured content of the component to be detected.

[0018] In the present invention, in step 4, the measured component sequence A of sampling period n n =[a n×1 ... an×i ...a n×I ], where a n×i is the measured content of the component i to be tested in sampling period n, and I is the total number of components to be tested.

[0019] In the present invention, in step 6, the measured component matrix B N×I =[A1 ... A n ... A N ] -1 = , timing drift sequence C n =[T1 ... T n ... T N ], T n is the sampling drift of sampling period n, the timing state sequence D n =[d1...d n ... d N ],d n is the state parameter of the batching system in sampling period n, and N is the total number of sampling periods.

[0020] In the present invention, in step 7, the component matrix E is modified N×I = , where the corrected content e of the component i to be tested in sampling period n is n×i = a n×i -g(T n ), if d n =d n-1 , g(T n ) is the intermediate interpolation function, if d n ≠d n-1 , g (T n ) is the forward interpolation function.

[0021] In the present invention, in step 7, the component prediction model consists of a global sub-model and a batch sub-model. The global sub-model is generated according to the corrected component matrix, the batch sub-matrix is ​​decomposed from the corrected component matrix according to the time series state sequence, and the batch sub-model is generated according to the batch sub-matrix.

[0022] In the present invention, the state parameters of the time series state sequence are traversed in reverse from the sampling period N until the state parameters meet d N-m =1 and d N-m-1 = 0, batch submatrix K N×I = , m is the number of traversals.

[0023] In the present invention, the first abnormal signal includes the type of the component to be tested that is outside the qualified interval, the actual measured content of the component to be tested, the current sampling cycle and the actual measured identifier, and the second abnormal signal includes the type of the component to be tested that is outside the qualified interval, the predicted content of the component to be tested, the current sampling cycle and the predicted identifier.

[0024] The implementation of a feed production process monitoring method based on ingredient prediction of the present invention has the following beneficial effects: the present invention sends a first abnormal signal according to whether the measured content of the component to be tested is within the qualified interval, and sends a second abnormal signal according to whether the predicted content is within the qualified interval, which can predict unqualified feed products in advance and avoid the waste of raw materials entering the processing system. In view of the delay problem of sampling during feed monitoring, the present invention records the sampling drift of each sampling, obtains a uniformly spaced corrected ingredient matrix after calibration, and provides accurate sample data for prediction. The present invention creates a time series state sequence in combination with the unloading state of the batching system, inserts batch marks in the corrected ingredient matrix, and then constructs a component prediction model based on this to improve the accuracy of monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 The figure is a schematic diagram of a feed production line;

[0026] Figure 2 A flow chart of a feed production process monitoring method based on ingredient prediction of the present invention;

[0027] Figure 3 It is a comparison diagram of two mass spectrometry data of the present invention;

[0028] Figure 4 A curve diagram of a response function of the present invention;

[0029] Figure 5 It is a time sequence state diagram of the batching system and the processing system of the present invention;

[0030] Figure 6 Schematic diagram of sampling drift of different sampling periods of the present invention;

[0031] Figure 7 A schematic diagram of constructing a component prediction model of the present invention;

[0032] Figure 8 Schematic diagram of the modified component matrix of the present invention. DETAILED DESCRIPTION

[0033] In order to more clearly understand the purpose, technical solutions and advantages of the present application, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0034] Complete feed is made by mixing multiple ingredients, and then packaged into bagged feed after conditioning and pelleting. Figure 1 , the corresponding feed production line includes a batching system, a processing system and a packaging system, and the processing system includes processes such as tempering, granulation and cooling. In order to avoid unqualified feed products, a mass spectrometry detection system is usually attached to the processing system to detect in real time whether heavy metals, toxins or pesticide residues are within the qualified range. There is an unstable time lag in the operation of the mass spectrometry detection system, that is, the actual sampling interval is uneven. At the same time, due to the periodicity of the working state of the batching system, there are large differences in the content of the components to be tested in different batches. Uneven sampling and batch changes make it difficult to predict the components to be tested. The present invention discloses a feed production process monitoring method based on component prediction, which first performs homogenization preprocessing on the sampling data, and then generates the predicted content of the components to be tested in the next sampling cycle in combination with the working state of the batching system, and finally completes the prediction of the components to be tested. Embodiment 1

[0035] like Figures 2 to 7 A feed production process monitoring method based on component prediction of the present invention comprises the following steps:

[0036] Step 1: The batching system periodically mixes at least two raw materials, and the processing system receives the mixed raw materials and produces a feed product, which includes a matrix and at least one component to be tested. The raw materials include at least one or more of corn bran, soybean meal, shell meal, and meat and bone meal, the matrix is ​​one or more of vitamins, proteins, or carbohydrates, and the component to be tested is one or more of heavy metals, toxins, or pesticides.

[0037] Step 2: Collect mass spectrometry data of feed products in the processing system at intervals of one sampling period, identify multiple signal peaks according to the mass spectrometry data, identify the signal peak of the matrix according to the calibrated content of the matrix ion, and calculate the sampling drift according to the acquisition time of the signal peak. The mass spectrometry detection system is, for example, a quadrupole mass spectrometer (QMS) or a time-of-flight mass spectrometer (TOF-MS). The mass spectrometry detection system periodically collects feed products in the processing system, ionizes the feed products into ions, and different ions fly at different speeds and hit the ion receiver. The signal intensity of the ion receiver at different times is collected, and finally the mass spectrometry data is generated according to the signal intensity at different acquisition times. In the process of continuous detection, the corresponding time difference of the mass spectrometry detection system gradually accumulates, and sampling drift occurs between adjacent sampling periods.

[0038] This embodiment provides a first mass spectrum generated by current mass spectrum data and a second mass spectrum generated by ideal mass spectrum data, such as Figure 3As shown. The higher the content, the greater the signal intensity. The matrix selects a component with a stable content, and the reference content is, for example, a known amount of matrix added. The signal intensity of the matrix is ​​determined based on the reference content of the matrix, and a signal peak with this signal intensity is searched in the first mass spectrum, and the collection time of the signal peak is used as the first collection time t1. The calibration collection time t2 is calculated based on the mass-to-charge ratio of the matrix ions. Specifically, in this embodiment, the flight time can be calculated based on the mass-to-charge ratio of the matrix ions. The flight time of the ions is proportional to the mass-to-charge ratio. The calibration collection time is then obtained by superimposing the flight time by the system excitation time, which will not be elaborated here. Figure 3 As shown, the sampling drift is calculated based on the first acquisition time t1 and the calibration acquisition time t2, that is, the sampling drift T of the sampling period n. n = t1- t2.

[0039] Step 3: Extract the signal peak of at least one component to be detected, calculate the mass-to-charge ratio of the component to be detected based on the acquisition time of the signal peak, and calculate the measured content of the component to be detected based on the signal intensity of the signal peak. The flight time of the corresponding ion is obtained by subtracting the excitation time of the current sampling cycle of the mass spectrometry detection system from the acquisition time of the signal peak of the component to be detected (without considering the working time of the mass spectrometry detection system), and then the mass-to-charge ratio of the corresponding ion is calculated from the flight time, thereby determining the molar mass and corresponding type of the component to be detected.

[0040] Calculate multiple response ratios based on the signal intensities of multiple standardized content components in the feed product. The response ratio is the ratio of the signal intensity to the total intensity. Fit the response ratios under multiple standardized contents into a smooth response curve. Figure 4 As shown in FIG. 1 , the response curve is the response function of the mass spectrometry detection system. The larger the response ratio, the higher the content. The response ratio of the component to be detected is generated according to the signal intensity of the signal peak of the component to be detected, and the response ratio is substituted into the response function to obtain the actual content of the component to be detected.

[0041] Step 4: If the measured content of at least one component to be tested is outside the qualified interval, the first abnormal signal is output and the task is terminated. Otherwise, a measured component sequence is generated from the measured contents of multiple components to be tested. The qualified interval can be the content standard of the component to be tested in the relevant industry. The maximum allowable amount of lead should not usually exceed 5 mg / kg. The qualified interval can be [0, 0.005‰], or a more stringent setting of [0, 0.001‰]. If the measured content of at least one component to be tested is outside the qualified interval, that is, the feed product produced this time is an unqualified product, the first abnormal signal is output immediately and the task is terminated. If the measured contents of all components to be tested are within the qualified interval, a measured component sequence is generated. Measured component sequence A for sampling period n n =[a n×1 ... a n×i ... a n×I ], where an×i is the measured content of the component i to be tested in sampling period n, and I is the total number of components to be tested.

[0042] Step 5: If the batching system is in the feeding state, the state parameter of the batching system is set to 1, otherwise the state parameter of the batching system is set to 0. Figure 5 First, the batching system enters the unloading state, and the processing system enters the processing state. Feed products have many processing steps, and the batching speed is greater than the processing speed. After completing the delivery of a batch of raw materials, the batching system enters the shutdown state, and the processing system maintains a continuous processing state. After the current batch of raw materials is used up, the batching system enters the unloading state again. The mass spectrometry detection system cooperates with the processing system to continuously detect the feed products until the processing system enters the shutdown state. The present invention records the state parameters of the current sampling cycle by judging whether the batching system is in the shutdown state during the current sampling cycle. The state parameters can be used to judge whether the current batch is in the process of updating the batch, which is convenient for considering the ingredient differences between different batches of feed products when constructing the ingredient prediction model later.

[0043] Step 6: Repeat steps 2 to 5 to generate a measured component matrix from the measured component sequences of multiple sampling periods, generate a timing drift sequence from the sampling drifts of multiple sampling periods, and generate a timing state sequence from the state parameters of multiple sampling periods. Measured component matrix B N×I =[A1 ... A n ... A N ] -1 = , N is the total number of sampling cycles. Timing drift sequence C n =[T1 ... T n ... T N ], T n is the sampling drift of sampling period n, the timing state sequence D n =[d1...d n ... d N ],d n is the state parameter of the batching system in sampling period n.

[0044] This embodiment discloses the measured content of lead in a complete feed, as shown in the following table.

[0045]

[0046] The same batch of raw materials has the same source, and the measured content of lead is similar, but the measured content of lead between different batches is different. The present invention trains a batch sub-model based on the measured content of the same batch, and trains a global sub-model based on the measured content of different batches, which can improve the accuracy of the model, as described in Example 3 and Example 4.

[0047] Step 7: Adjust the measured component matrix to a modified component matrix according to the time series drift sequence, and construct a component prediction model based on the modified component matrix and the time series state sequence. Figure 6 Due to the existence of sampling drift, the production line has non-uniform sampling phenomenon, that is, the actual sampling interval in the measured component matrix is ​​not equal to the preset sampling interval T0. Therefore, in order to ensure the accuracy of subsequent data prediction, it is necessary to homogenize the measured component matrix and obtain the corrected component matrix under uniform sampling. Corrected component matrix E N×I = , the corrected content e of the component i to be tested in sampling period n n×i = a n×i -g(T n ), if d n =d n-1 , g(T n ) is the intermediate interpolation function, if d n ≠d n-1 , g(T n ) is a forward interpolation function. For details, refer to the second embodiment.

[0048] Preferably, the component prediction model consists of a global sub-model and a batch sub-model, such as Figure 7 As shown. The global sub-model is used to describe the content relationship of the components to be tested in all batches of raw materials, and the batch sub-model is used to describe the content relationship of the components to be tested in the current batch of raw materials. A global sub-model is generated according to the corrected component matrix, and a batch sub-matrix is ​​decomposed from the corrected component matrix according to the time series state sequence, and a batch sub-model is generated according to the batch sub-matrix. The output values ​​of the global sub-model and the batch sub-model are combined into the final predicted content. By adopting the model splitting method of this embodiment, the changes in raw materials of different batches can be taken into account, thereby improving the accuracy of the prediction.

[0049] Step 8: Generate a predicted component sequence for the next sampling period based on the component prediction model, and extract the predicted content of at least one component to be tested in the predicted component sequence. Depending on the type of the component prediction model, the corresponding input parameters can be input into the component prediction model to obtain the predicted component sequence. The predicted component sequence consists of the predicted content of each component to be tested in the N+1th sampling period.

[0050] Step 9: If the predicted content of at least one component to be tested is outside the qualified interval, output a second abnormal signal and end the task, otherwise return to step 2. After returning to step 2, collect the measured content of N+1 sampling cycles and update the measured component matrix. The present invention does not limit the output content of the first abnormal signal and the second abnormal signal. In this embodiment, the first abnormal signal includes the type of component to be tested outside the qualified interval, the measured content of the component to be tested, the current sampling cycle and the measured identifier, and the second abnormal signal includes the type of component to be tested outside the qualified interval, the predicted content of the component to be tested, the current sampling cycle and the predicted identifier. Embodiment 2

[0051] This embodiment further describes in detail the data processing method in step 7 of the first embodiment.

[0052] Corrected content e of component i to be tested n×i = a n×i -g(T n ). Due to batch variations, the component variable function g(T n ) may mutate, and g(T n ) algorithm. If d n =d n-1 , sampling period n and n-1 are in the same batch, and g(T n ). At this time g(T n ) is the intermediate interpolation function, g(T n )= If d n ≠d n-1 , sampling period n and n-1 are in different batches, the interpolation data is discontinuous with the previous data, and g(T) is generated according to the measured content of sampling period n+1 n ). At this time g(T n ) is the forward interpolation function, g(T n )= The present invention adopts Newton interpolation method to interpolate forward, which can be applied to the situation where the previous data is discontinuous. The correction content of the sampling period N is not adjusted.

[0053] like Figure 8 As shown, the batch submatrix is ​​a partial sequence of the modified component matrix. The state parameters are traversed in reverse from the sampling period N until the state parameters meet d N-m =1, and d N-m-1 =0, at which point the sampling period Nm is the first sampling period of the current batch of raw materials, and m is the number of traversals. Decompose the batch submatrix from the corrected component matrix, which contains the measured content from sampling period Nm to N. The batch submatrix K N×I = , the batch sub-matrix can be used as sample data for the batch sub-model. Embodiment 3

[0054] This embodiment discloses a preferred method for constructing a component prediction model. This embodiment uses a cubic polynomial function to construct a global sub-model and a batch sub-model of the component i to be detected.

[0055] First, construct a cubic polynomial function of the global submodel. The cubic polynomial function of the interval [n, n+1] is S n (x), S n (x)=α n +β n (xn)+γ n (xn) 2 +δ n (xn) 3 , x is the independent variable of the polynomial.

[0056] Then, determine the conditions of the cubic polynomial function. The conditions of the cubic polynomial function are not limited to one or more of the following conditions. Function continuity condition: S n (n)=e n×i , S n (n+1)=e (n+1)×i ; First-order derivative continuity condition: S n '(n+1)=S n+1 '(n); Second-order derivative continuity condition: S n ''(n+1)=S n+1 ''(n); Quadratic coefficient condition: γ n =S n ''(n) / 2; boundary condition: S1''(1)=S N (N)''=0.

[0057] Then solve the coefficients of the cubic polynomial function. Starting from n=1, substitute the measured content of the correction component matrix into the cubic polynomial function in turn, solve the coefficients of the above n groups of cubic polynomial functions, and finally obtain the cubic polynomial function S when n=N N (x), S N (x) is stored as a global submodel.

[0058] In the same way, construct a cubic polynomial function F in the interval [Nm, N-m+1] N-m (x), and then substitute the measured content of the batch sub-matrices in turn, and finally obtain the cubic polynomial function F when n=N N (x), F N (x) are stored as batch sub-models.

[0059] N+1 is the input parameter, x=N+1 is substituted into the global sub-model and the batch sub-model, and the output values ​​are combined to obtain the predicted content, and then the predicted content of the I group of components to be tested is used to form a predicted component sequence. For example, the predicted content e (N+1)×i =[S N (N+1)+F N (N+1)] / 2. In another embodiment, S N (N+1) and F N (N+1) Set different merging weights and then calculate the predicted content. Embodiment 4

[0060] This embodiment discloses another preferred method for constructing a component prediction model. In this embodiment, the global sub-model and batch sub-model of the component i to be tested are both recurrent neural network models, and the recurrent neural network model includes an input layer, a hidden layer, and an output layer. The generation method of the global sub-model is as follows.

[0061] Step 100: Use the modified component matrix as sample data to generate a global sub-model. Input the first N-1 rows of the modified component matrix into the recurrent neural network model in sequence, and output the predicted sequence Y through forward propagation. n ={y n×1 ,…,y n×i ,…, y n×I}.

[0062] Step 200: If the current prediction sequence Y n The measured component sequence A n If the error is within the expected range, stop training and use the current recurrent neural network model as the global sub-model. Otherwise, go to step 300 and continue the recurrent training.

[0063] Step 300: Define the loss function. In this embodiment, the total loss function is defined based on the mean square error as , where the loss function L of the nth sampling period is n =(e n×i -y n×i ) 2 / 2.

[0064] Step 400: Substitute the predicted component sequence and the measured component sequence of each sampling period into the loss function of each sampling period, and calculate the loss function L of each sampling period respectively. n Then, the recurrent neural network model is iteratively updated according to the loss function, and the process returns to step 100.

[0065] The batch sub-matrix is ​​used as sample data to generate the batch sub-model. First, the Nm-1 to N-1 rows of the batch sub-matrix are substituted into the recurrent neural network model, and after the prediction sequence is generated, it is compared with the Nm to N rows of measured component sequences, and finally the recurrent neural network model is updated according to the comparison results.

[0066] The measured content of sampling period N is the input parameter. The measured content of the component i to be tested in sampling period N is input into the global sub-model and the batch sub-model to obtain two sets of candidate sequences. The two sets of candidate sequences are then averaged to obtain the predicted content of the component i to be tested.

[0067] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for monitoring feed production process based on component prediction, characterized in that: The following steps are involved: Step 1: a batching system periodically mixes at least two raw materials, and a processing system receives the mixed raw materials and produces a feed product, wherein the feed product includes a matrix and at least one component to be tested; Step 2: Collect mass spectrum data of feed products in the processing system at every sampling period, identify multiple signal peaks according to the mass spectrum data, identify the signal peak of the matrix according to the calibrated content of the matrix ions, and calculate the sampling drift according to the acquisition time of the signal peak; Step 3: extracting a signal peak of at least one component to be detected, calculating the mass-to-charge ratio of the component to be detected according to the acquisition time of the signal peak, and calculating the measured content of the component to be detected according to the signal intensity of the signal peak; Step 4: If the measured content of at least one component to be tested is outside the qualified range, a first abnormal signal is output and the task is terminated; otherwise, a measured component sequence is generated based on the measured contents of multiple components to be tested; Step 5: If the batching system is in the feeding state, the state parameter of the batching system is set to 1, otherwise the state parameter of the batching system is set to 0; Step 6: Repeat steps 2 to 5, generate a measured component matrix from the measured component sequences of multiple sampling periods, generate a timing drift sequence from the sampling drifts of multiple sampling periods, and generate a timing state sequence from the state parameters of multiple sampling periods; Step 7: Adjust the measured component matrix to a modified component matrix according to the time series drift sequence, and construct a component prediction model according to the modified component matrix and the time series state sequence; Step 8: Generate a predicted component sequence for the next sampling period according to the component prediction model, and extract the predicted content of at least one component to be tested in the predicted component sequence; Step 9: If the predicted content of at least one component to be tested is outside the qualified range, output a second abnormal signal and end the task; otherwise, return to step 2. The measured component sequence A of sampling period n n =[a n×1 ... a n×i ... a n×I ], where a n×i is the measured content of the component i to be tested in sampling period n, I is the total number of components to be tested, Measured component matrix B N×I =[A1 ... A n ... A N ] -1 = , timing drift sequence C n =[T1 ... T n ... T N ], T n is the sampling drift of sampling period n, the timing state sequence D n =[d1...d n ...d N ],d n is the state parameter of the batching system in sampling period n, N is the total number of sampling periods, In step 7, the component matrix E is modified N×I = , where the corrected content e of the component i to be tested in sampling period n is n×i = a n×i -g(T n ), if d n =d n-1 , g(T n ) is the intermediate interpolation function, if d n ≠d n-1 , g (T n ) is the forward interpolation function, The component prediction model consists of a global sub-model and a batch sub-model. The global sub-model is generated according to the modified component matrix. The batch sub-matrix is ​​decomposed from the modified component matrix according to the time series state sequence. The batch sub-model is generated according to the batch sub-matrix. The global sub-model is used to describe the content relationship of the components to be tested in all batches of raw materials, and the batch sub-model is used to describe the content relationship of the components to be tested in the current batch of raw materials. The output values ​​of the global sub-model and the batch sub-model are combined into the final predicted content.

2. The feed production process monitoring method based on component prediction according to claim 1, characterized in that: In step 1, the raw materials include at least one or more of corn bran, soybean meal, shell meal, and meat and bone meal, the matrix is ​​one or more of vitamins, proteins, or carbohydrates, and the components to be tested are one or more of heavy metals, toxins, or pesticides.

3. The feed production process monitoring method based on component prediction according to claim 1, characterized in that: In step 2, the signal intensity of the matrix is ​​determined according to the reference content of the matrix, a signal peak with the signal intensity is found, the acquisition time of the signal peak is used as the first acquisition time, the calibration acquisition time is calculated according to the mass-to-charge ratio of the matrix ions, and the sampling drift is calculated according to the first acquisition time and the calibration acquisition time.

4. The feed production process monitoring method based on component prediction according to claim 1, characterized in that: In step 3, a response ratio of the component to be detected is generated according to the signal intensity of the signal peak of the component to be detected, a response function is fitted, and the response ratio is substituted into the response function to obtain the actual measured content of the component to be detected.

5. The feed production process monitoring method based on component prediction according to claim 1, characterized in that: From the sampling period N, the state parameters of the time series state sequence are traversed in reverse order until the state parameters meet d N-m =1 and d N-m-1 = 0, batch submatrix K N×I = , m is the number of traversals.

6. The feed production process monitoring method based on component prediction according to claim 1, characterized in that: The first abnormal signal includes the type of component to be tested that is outside the qualified interval, the measured content of the component to be tested, the current sampling cycle and the measured identifier; the second abnormal signal includes the type of component to be tested that is outside the qualified interval, the predicted content of the component to be tested, the current sampling cycle and the predicted identifier.

Images