Abnormal detection method and system for cutting precision of granulator

Abstract

The invention relates to the technical field of granulator anomaly detection, and discloses a granulator cutting precision anomaly detection method and system. According to the method, a vibration signal and a three-phase load current signal during operation of the granulator are collected in real time, an impact characteristic spectrum and a current distortion factor are extracted through time-frequency domain transformation and harmonic decomposition, a pre-trained deep learning model is input for time sequence characteristic fusion, and a cutting state vector is generated. And judging that the cutting precision is abnormal based on dynamic threshold comparison, triggering an alarm, and marking raw material batch information associated with an abnormal time period. According to the method, the accuracy and real-time performance of cutting precision anomaly detection are remarkably improved, the method adapts to dynamic factors such as raw material characteristic changes and tool abrasion, unqualified products are reduced, and the production cost and market risks are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of abnormal detection of dicing machine, and particularly relates to an abnormal detection method and system for cutting precision of dicing machine. BACKGROUND

[0002] In the production process of bonding wires, the stability of cutting precision is crucial for product quality. However, the traditional detection method often fails to capture the tiny abnormalities in the cutting process in real time and accurately, resulting in low detection precision and poor adaptability to complex working conditions. Especially in the face of dynamic factors such as changes in raw material characteristics and tool wear, the traditional method is even more inadequate. These problems not only affect production efficiency, but also may lead to the production of a large number of unqualified products, increasing the production cost and market risk of enterprises. Therefore, developing an efficient and accurate abnormal detection method and system for cutting precision of dicing machine has become an urgent need in current industrial production. SUMMARY

[0003] In view of the above technical deficiencies, the purpose of the present application is to provide an abnormal detection method and system for cutting precision of dicing machine, which solves the problem of low detection precision of cutting precision of dicing machine in the prior art.

[0004] To solve the above technical problems, the present application adopts the following technical solutions: In a first aspect, the present application provides an abnormal detection method for cutting precision of dicing machine, which comprises: Real-time acquisition of multi-source signals during operation of the dicing machine, including vibration signals of the cutting device and three-phase load current signals of the driving motor; Time-frequency domain transformation of the vibration signals to extract impact feature spectrum, and harmonic decomposition of the three-phase load current signals to generate current distortion factor; Inputting the impact feature spectrum and the current distortion factor into a pre-trained deep learning model to generate a cutting state vector through time sequence feature fusion; Determining cutting precision abnormality according to the comparison result of the cutting state vector and a dynamic threshold, wherein the dynamic threshold is updated based on historical normal working condition data through an adaptive algorithm; When an abnormality is detected, triggering an alarm and marking the raw material batch information associated with the abnormal period.

[0005] Preferably, in a possible implementation of the first aspect, the time-frequency domain transformation specifically comprises: Empirical mode decomposition of the vibration signals to obtain intrinsic mode function components; Fourier transform of the first component to obtain instantaneous frequency and instantaneous amplitude at a certain time ; Constructing impact signature spectrum :

[0006] where is the Dirac function, is the frequency axis variable, is the time variable, is the total number of eigenmode function components.

[0007] Preferably, in a possible implementation form of the first aspect, the harmonic decomposition specifically comprises: performing fast Fourier transform on the three-phase load current signal to extract the fundamental component amplitude and the harmonic component amplitudes , ; current distortion factor is:

[0008] where is the preset upper limit of harmonic order, when , it is determined that the current waveform is abnormally distorted, and the cutting resistance is suddenly changed.

[0009] Preferably, in a possible implementation form of the first aspect, the pre-trained deep learning model is a dual-channel time convolution network, comprising: the first channel inputs the time sequence of the impact signature spectrum, extracts the tool wear time sequence feature through causal convolution; the second channel inputs the current distortion factor sequence, and adopts dilated convolution to capture long-period load features; the dual-channel outputs are input into an attention fusion layer after residual connection.

[0010] Preferably, in a possible implementation form of the first aspect, the attention fusion layer performs a gated cross-attention mechanism: the first channel output feature matrix is , the second channel output is , and the fusion weight matrix W is calculated by the following formula:

[0011] where , is a trainable parameter matrix, is a softmax function, denotes Hadamard product, is an activation function, is a time step, is a feature dimension, is a real number matrix.

[0012] Preferably, in a possible implementation form of the first aspect, the temporal feature fusion process specifically comprises: performing gated fusion on the first channel output feature matrix and the second channel output feature matrix using a fusion weight matrix W to generate a fused feature matrix :

[0013] wherein represents a Hadamard product, is an all-1 matrix, is the fused feature matrix; input feature compression layer generates a sequence of cut state vectors :

[0014] wherein is a compression layer weight matrix, is a bias vector, is a cut state vector dimension, is an output sequence.

[0015] Preferably, in a possible implementation form of the first aspect, the dynamic threshold comparison specifically comprises: cut state vector input abnormality scorer calculates Mahalanobis distance , when is abnormal; wherein is a historical normal state vector mean, represents a cut state vector at the moment, is a transposition operation, is a covariance matrix, is a dynamic threshold at the moment.

[0016] Preferably, in a possible implementation form of the first aspect, the dynamic threshold is updated by an adaptive algorithm: take the latest hours of normal operating data, calculate the sliding mean and standard deviation of the state vector, and the dynamic threshold update formula is:

[0017] wherein ​a base threshold coefficient, a time decay factor, a system operating hours.

[0018] Preferably, in a possible implementation of the first aspect, the base threshold coefficient Dynamic adjustment according to raw material characteristics:

[0019] wherein , an adjustment coefficient, a current raw material batch hardness value, a current raw material batch viscosity value, , a standard raw material parameter reference value.

[0020] In a second aspect, the present application provides an abnormality detection system for cutting precision of a dicer, comprising: a signal acquisition module, which acquires multiple source signals in real time when the dicer is running, including vibration signals of a cutting device and three-phase load current signals of a driving motor; a feature extraction module, which extracts impact feature spectrum by time-frequency domain transformation on the vibration signals, and generates current distortion factors by harmonic decomposition on the three-phase load current signals; a feature fusion module, which inputs the impact feature spectrum and the current distortion factors into a pre-trained deep learning model, and generates a cutting state vector by time sequence feature fusion; an abnormality determination module, which determines cutting precision abnormality according to a comparison result of the cutting state vector and a dynamic threshold, wherein the dynamic threshold is updated by an adaptive algorithm based on historical normal working condition data; an alarm triggering module, which triggers an alarm and marks raw material batch information associated with an abnormal period when an abnormality is detected.

[0021] The present application has the beneficial effects that: by acquiring multiple source signals in real time and utilizing machine learning model for feature fusion and abnormality determination, the accuracy and real-time performance of cutting precision abnormality detection are significantly improved.

[0022] The method and system are highly adaptable, can effectively cope with dynamic factors such as raw material characteristic changes and tool wear, reduce the production of unqualified products, and reduce production costs and market risks.

[0023] At the same time, by dynamically updating the threshold and associating the raw material batch information, the reliability and traceability of the detection are further improved, providing a strong guarantee for industrial production. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0025] Figure 1 A flow chart of the abnormality detection method of the cutting precision of the dicer is provided for the present application.

[0026] Figure 2 A system structure diagram of the abnormality detection system of the cutting precision of the dicer is provided for the present application.

[0027] Legend: 1-signal acquisition module, 2-feature extraction module, 3-feature fusion module, 4-abnormality determination module, 5-alarm triggering module. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0029] Embodiment one: as shown, the present application provides an abnormality detection method of the cutting precision of the dicer, comprising: Figure 1 Real-time acquisition of multi-source signals of the dicer during operation, including vibration signals of the cutting device and three-phase load current signals of the driving motor. In the present embodiment, the acquisition of multi-source electrical signals is realized by sensors deployed on the dicer. Specifically, a high-sensitivity, wide-frequency-response piezoelectric acceleration sensor (PCB 352C33 type is used in the present embodiment) is installed on the cutting device near the rotating plane of the cutter, which is used to capture the vibration impact signals generated during the contact and cutting process of the cutter and the material in real time. The signals reflect the states such as cutter wear, edge collapse or sudden change of material hardness. At the same time, a high-precision closed-loop Hall current sensor (LEM LF 510-S type is used in the present embodiment) is connected in series on the driving motor power line, which is used to non-invasively and synchronously measure the three-phase load current, and to obtain the current waveform data reflecting the motor load change, torque fluctuation and potential electrical faults (such as open phase, imbalance) in real time.

[0030]

[0031] ​To eliminate the strong electromagnetic interference and ensure the signal fidelity, the sensor signals are connected to the data acquisition card (NI 9223 in this embodiment) with synchronous sampling function through shielded cable. The acquisition card has a synchronous sampling rate of at least 10 kHz and a resolution of 16 bits, and is configured with an anti-aliasing filter. The entire acquisition process is continuously carried out under the scheduling of the control system PLC, with a sampling batch every 100 milliseconds.

[0032] The vibration signal is subjected to time-frequency domain transformation to extract the impact feature spectrum, and the three-phase load current signal is subjected to harmonic decomposition to generate the current distortion factor.

[0033] In this embodiment, the vibration signal processing adopts an adaptive time-frequency analysis method: first, the vibration signal (sampling rate 10 kHz) collected by the piezoelectric sensor is subjected to empirical mode decomposition, and the screening stop condition is set to be less than 0.3 for the standard deviation of continuous two screenings, to obtain 5-8 intrinsic mode function components. For the 2-4 order IMF components (located in the 1.5-8 kHz frequency band) reflecting the impact characteristics of the tool, the transform is applied to each component with a window length of 50 points to calculate the time instant frequency and the instantaneous amplitude , and the mirror extension method is used to eliminate the end effect. When constructing the impact feature spectrum, the time axis is divided into 10 ms windows, and the impact feature spectrum of a 256x1000 matrix is generated in the three-dimensional time-frequency space according to the formula , where is the Dirac function, is the frequency axis variable, is the time variable, is the total number of intrinsic mode function components, and the matrix represents the transient distribution characteristics of the tool-material interaction energy in the time-frequency domain.

[0034] In terms of current signal processing, the synchronous three-phase current (sampling rate 5 kHz) collected by the Hall sensor is subjected to three-phase decoupling processing: first, the three-phase current is converted to a two-phase orthogonal system through transform to eliminate the interference of fundamental frequency fluctuation. The fast Fourier transform of the window function (Blackman window, window length 100 ms) is applied to each phase current to extract the fundamental component amplitude and the 2-19 harmonic component amplitude . The current distortion factor is calculated in real time according to the formula , and when lasts more than , it is associated with the cutting resistance sudden change condition (such as material clumping or tool blade collapse). A sliding window update mechanism is adopted: every Update the Fast Fourier Transform calculation results only if there are 5 consecutive windows. The flag is triggered only when the value exceeds the threshold.

[0035] A dual verification mechanism is implemented in the signal processing stage: envelope entropy detection is performed on the vibration signal (resampling is triggered when the entropy value exceeds 0.7), and three-phase unbalance monitoring is performed on the current signal (signal compensation is initiated when the unbalance is >10%). The processed impact characteristic spectrum matrix and distortion factor sequence are aligned with timestamps to form a synchronous characteristic stream.

[0036] The impact feature spectrum and current distortion factor are input into a pre-trained deep learning model, and a cutting state vector is generated by fusing temporal features.

[0037] In this embodiment, the feature fusion module employs a dual-channel temporal convolutional network architecture to perform multimodal temporal feature fusion. The network input layer is designed as a dual-path parallel structure: the first channel receives the preprocessed time series of the impact feature spectrum, which is of dimensionality... The matrix (frequency point × time step) is reorganized into a 256-dimensional feature vector of 1000 consecutive time steps through time slices; the second channel input current distortion factor sequence, this sequence is... The scalar value of the sampling interval is integrated through a sliding window. step A 3D time series. The dual-channel input data is aligned with timestamps to ensure time-domain consistency.

[0038] The first channel's temporal processing employs a 5-layer stacked causal convolutional structure. Each layer has 64 convolutional kernels, with a kernel width set to 7 time steps and a fixed stride of 1. To achieve progressive extraction of temporal features, a batch normalization layer is inserted after the first convolutional layer. Activation function, output dimension remains the same. (Time step × Feature dimension). The subsequent four convolutional layers all use a residual connection structure: the output of each layer is connected to the input via... After dimension matching, the convolutional layers perform element-wise addition to alleviate the gradient vanishing problem. Specifically, the third convolutional layer introduces dilated convolutions with a dilation rate of 2, expanding the temporal receptive field to 35 time steps (covering...). (Operating conditions), capturing the slowly varying characteristics of progressive tool wear. The final first channel output tensor... Its row vectors represent the high-dimensional features of the tool impact state at each time step.

[0039] The second channel addresses the long-period characteristics of the current distortion factor by constructing a four-layer dilated convolution module. The initial layer uses standard 1D convolution (kernel width 5, stride 1) to expand the single-dimensional input to a 64-dimensional feature space. The subsequent three layers are set with an exponentially increasing dilation rate. , the kernel width is unified as 3. This design makes the receptive field of the top-level convolution expand to 93 time steps , effectively associating the minute-level fluctuation pattern of the load current. After each convolution layer, an exponential linear unit activation is followed, and a gating mechanism is introduced in the skip connection: the gating coefficient of the interval is generated by a sigmoid transformation , dynamically modulating the fusion ratio of the original input and the convolution output. The final output tensor , whose time dimension is synchronized with the first channel.

[0040] The feature fusion core process is completed in the attention fusion layer. The input tensor 、 is projected by linear transformation: to , and Q is transformed to . The gated cross-attention calculation is performed: first, the hyperbolic tangent nonlinear transformation is applied to , and the activation function is applied to to generate a gating mask. The two are combined by Hadamard product to realize feature interaction: , where represents the normalization along the time dimension, is the activation function. This mechanism makes the fusion weight matrix have a dual characteristic: in the spatial dimension, 64 feature channels independently calculate attention weights; in the time dimension, each time step automatically allocates , the contribution ratio of the features.

[0041] The temporal feature fusion performs weighted synthesis: a full-1 matrix is constructed, and the fusion feature matrix is calculated according to the formula . Where represents element-level multiplication, which is equivalent to constructing a dynamic feature selector: when an element in W is close to 1, the corresponding feature in the P channel is preferentially selected; when it is close to 0, the Q channel feature is adopted. The fusion output is input into the feature compression layer for dimension reduction: this layer is a fully connected network with weight matrix and bias vector . Through linear transformation , an intermediate tensor is generated, and then the activation function is applied to output the cutting state vector sequence . The 8-dimensional vector at each time step represents: the first three dimensions represent the mechanical state such as tool wear and impact strength; the middle two dimensions correspond to the motor load fluctuation; and the last three dimensions reflect the material-tool interaction parameters.

[0042] Model training adopts historical three-month normal operating data, and the data set contains 12000 groups of synchronous samples. The training strategy adopts a five-step method: firstly, the double-channel is pre-trained for 50 rounds, and the loss function is set as mean square error; then the convolution layer parameters are frozen, and the attention fusion layer is trained for 30 rounds; then the whole network is fine-tuned for 100 rounds in an end-to-end manner, and the cosine annealing learning rate scheduling with warm restart is adopted (initial value 0.01, period 20 rounds); the early stopping strategy of the validation set is set to no improvement for 10 consecutive rounds; finally, the 8-bit fixed-point quantization is performed on the embedded system deployment of the final model, and the inference delay is controlled within .

[0043] According to the comparison result of the cutting state vector and the dynamic threshold, the cutting precision abnormality is determined, wherein the dynamic threshold is updated based on the historical normal operating data by an adaptive algorithm.

[0044] In this embodiment, the abnormality determination core mechanism adopts a dynamic threshold decision model based on Mahalanobis distance. Specifically, the system receives the cutting state vector sequence output by the feature fusion module in real time , and performs abnormal score calculation on the state vector of each time step . The scorer first calculates the Mahalanobis distance between and the historical normal state distribution: , wherein represents the mean vector of the state vector under the historical normal operating condition, is the corresponding covariance matrix, is the transpose operation. The distance measure is essentially the normalized deviation of the state vector under the multivariate Gaussian distribution, which can effectively eliminate the influence of the dimension difference and the correlation between different features. To ensure numerical stability, a regularization term is added when calculating the covariance matrix , and the form ( , is the unit matrix), to avoid failure of matrix inversion due to singularity. When the real-time calculated Mahalanobis distance exceeds the current dynamic threshold , the system determines that the cutting precision is abnormal at this moment, and triggers the abnormal flag.

[0045] The update of the dynamic threshold adopts an adaptive algorithm based on a sliding window, and its core formula is: . The formula consists of three parts: the reference term represents the central position of the state vector in the recent time window, the dispersion term characterizes the fluctuation range of the state vector, and the time decay term This is used for dynamically adjusting sensitivity. In practice, the system maintains a continuously updated pool of normal operating condition data, retaining only the Mahalanobis distance. Less than the current threshold The state vectors (i.e., data not identified as abnormal). Based on the normal data from the most recent N hours (N=24 in this example), the moving average of all state vectors within this window is calculated. and sliding standard deviation .in Calculated using an exponentially weighted moving average: , This is the current normal vector. Then an unbiased estimate of the standard deviation is used: , where k is the number of valid samples within the window. Time decay factor. (This embodiment takes) ) and system cumulative running hours Composition This feature allows the threshold to rise naturally over time, adapting to the slow degradation of equipment performance and avoiding false alarms caused by excessive sensitivity in the early stages.

[0046] Base threshold coefficient Considering the dynamic influence of raw material properties, the calculation formula is as follows: .in and The preset adjustment coefficient (in this embodiment, it is taken as...) , ), This indicates the measured Rockwell hardness value for the current batch of raw materials. Indicates the viscosity value of the raw material. and The reference values ​​for standard raw material parameters (as set in this embodiment) , When the raw material hardness Higher than the benchmark or viscosity Greater than hour, The value increases linearly, leading to a dynamic threshold. The system's criteria for anomaly detection are relaxed by raising the threshold. This is because harder or more viscous materials naturally generate greater mechanical impact and current fluctuations during cutting, making it prone to false alarms if standard thresholds are still used. Conversely, for softer or more fluid materials (…), the thresholds are lowered. or ), As the value decreases, the system's detection sensitivity increases. Function ensures Negative values ​​will not occur, maintaining the basic effectiveness of the threshold. Raw material parameters. and Real-time acquisition through enterprise MES system, bound to the current processing batch, automatically updated when the raw material batch is replaced.

[0047] System initialization phase, And Based on the normal cutting data of the equipment for 7 consecutive days after factory debugging, it is obtained offline, which ensures the coverage of typical raw materials and standard working conditions. In the process of continuous operation, the dynamic threshold update period is set to execute once every 15 minutes. Each update contains four steps: first, extract the valid normal working condition data of the last 24 hours (about 1440 groups Vector) from the rolling buffer; second, calculate the And of this window; then calculate the , value of the current raw material from the MES ; finally, calculate according to the current running hours . At the same time, the long-term statistics And are updated incrementally every 24 hours: , , where is the number of old data samples, is the number of new normal samples added on the day, And are the statistics of the data on the day. This incremental update mechanism enables the system to slowly adapt to the long-term performance drift of the equipment, such as the gradual wear of the tool mechanism, while the short-term dynamic threshold focuses on responding to sudden changes in working conditions such as raw material switching. To prevent data pollution, the normal working condition data pool has a strict access mechanism: only the data of the period when the alarm is not triggered for 1 hour and the operation log is marked as "standard mode" can be included in the statistics. In addition, the system has a consistency checking module, which automatically freezes the threshold update and triggers a manual review request when it detects that mutation exceeds 50% or value fluctuation is greater than 2.0.

[0048] When an anomaly is detected, an alarm is triggered and the raw material batch information associated with the abnormal period is marked.

[0049] In this embodiment, the alarm triggering module and the anomaly determination module are in real-time linkage, and when the dynamic threshold comparison result determines an anomaly for 3 consecutive time steps (i.e. ), a multi-level alarm protocol is activated. First, send a standard alarm signal to the control room main console through the industrial bus to drive the sound and light alarm to emit red stroboscopic and beep; at the same time, send structured alarm information (including abnormal timestamp, Mahalanobis distance score , suspicious reason classification code) to the workshop MES system. The raw material batch association mechanism is achieved by real-time polling of the MES production order database: the system automatically intercepts the complete time period from 10 seconds before the start of the abnormal window to 10 seconds after the end, indexes the device code and time interval, extracts the batch number, material code, process parameters (including hardness , viscosity measured value) and other key information of the current processing raw material from the batch tracking table of the MES, and generates an encrypted abnormal event record.

[0050] The record is packaged in JSON format, containing the original vibration signal segment, current waveform segment, and corresponding impact feature spectrum matrix, current distortion factor and other diagnostic data storage paths, and is persistently stored in the time series database while synchronously writing to the abnormal trace module of the quality management system. To reduce false positives, the system sets a secondary confirmation rule: only when a single abnormality lasts more than or the cumulative abnormality duration within 10 minutes accounts for more than 15%, will an alarm notification be pushed to the mobile terminal, and a red time axis marker covering the abnormal period will be automatically generated on the SCADA interface. The operator can view the raw material quality inspection report and cutting process parameter history curve of the associated batch by clicking on the marker.

[0051] Embodiment two: as shown in Figure 2 , the present application provides an abnormal detection system for cutting precision of a dicer, comprising: A signal acquisition module 1 acquires multiple source signals in real time during operation of the dicer, including vibration signals of the cutting device and three-phase load current signals of the driving motor.

[0052] A feature extraction module 2 extracts impact feature spectrum by performing time-frequency domain transformation on the vibration signals, and generates current distortion factors by harmonic decomposition of the three-phase load current signals.

[0053] A feature fusion module 3 inputs the impact feature spectrum and current distortion factors into a pre-trained deep learning model to generate a cutting state vector through time series feature fusion.

[0054] An abnormality determination module 4 determines cutting precision abnormalities based on the comparison results of the cutting state vector and the dynamic threshold, wherein the dynamic threshold is updated based on historical normal operating data through an adaptive algorithm.

[0055] An alarm triggering module 5 triggers an alarm and marks the raw material batch information associated with the abnormal period when an abnormality is detected.

[0056] It will be apparent to those skilled in the art that various modifications and variations can be made to the present application without departing from the spirit or scope of the application. Thus, it is intended that the present application cover modifications and variations of this application provided they come within the scope of the appended claims and their equivalents.

Claims

1. A method for detecting abnormalities in the cutting accuracy of a pelletizer, characterized in that, The method includes: Real-time acquisition of multi-source signals during pelletizer operation, including vibration signals from the cutting device and three-phase load current signals from the drive motor; The vibration signal is subjected to time-frequency domain transformation to extract the impact feature spectrum, and the three-phase load current signal is subjected to harmonic decomposition to generate current distortion factor. The impact feature spectrum and current distortion factor are input into a pre-trained deep learning model, and a cutting state vector is generated by time-series feature fusion. The cutting accuracy is determined based on the comparison result between the cutting state vector and the dynamic threshold, wherein the dynamic threshold is updated by an adaptive algorithm based on historical normal working condition data. When an anomaly is detected, an alarm is triggered and the raw material batch information associated with the anomaly period is marked.

2. The method for detecting abnormalities in the cutting accuracy of a pelletizer as described in claim 1, characterized in that, The time-frequency domain transformation specifically includes: Empirical mode decomposition is performed on the vibration signal to obtain the intrinsic mode function components; For the Each component is processed Transformation to obtain instantaneous frequency and instantaneous amplitude ; Constructing impact characteristic spectrum : in For the Dirac function, For frequency axis variables, For time variables, This represents the total number of intrinsic mode function components.

3. The method for detecting abnormalities in the cutting accuracy of a pelletizer as described in claim 1, characterized in that, The harmonic decomposition specifically includes: Perform a Fast Fourier Transform on the three-phase load current signal to extract the amplitude of the fundamental component. and amplitude of each harmonic component , ; Current distortion factor The formula is: in To preset the upper limit of harmonic order, when When the current waveform is abnormally distorted, it is associated with a sudden change in cutting resistance.

4. The method for detecting abnormalities in the cutting accuracy of a pelletizer as described in claim 1, characterized in that, The pre-trained deep learning model is a dual-channel temporal convolutional network, including: The first channel inputs the time series of the impact feature spectrum, and extracts the tool wear time series features through causal convolution; The second channel input current distortion factor sequence is used to capture long-cycle load characteristics using dilated convolution. The dual-channel output is fed into the attention fusion layer after being connected via residuals.

5. The method for detecting abnormalities in the cutting accuracy of a pelletizer as described in claim 4, characterized in that, The attention fusion layer executes a gated cross-attention mechanism: The first channel output feature matrix is The second channel output is The fusion weight matrix W is calculated using the following formula: in , For a trainable parameter matrix, For the softmax function, It represents the Hadamah accumulation. For activation function, For time step, For feature dimension, It is a real matrix.

6. The method for detecting abnormalities in the cutting accuracy of a pelletizer as described in claim 5, characterized in that, The temporal feature fusion process specifically includes: The first channel output feature matrix is ​​obtained by using the fusion weight matrix W. Second channel output feature matrix Perform gated fusion to generate a fusion feature matrix. : in It represents the Hadamah accumulation. It is a matrix of all ones. For fusion feature matrix; Will Input feature compression layer generates cut state vector sequence : in Here is the weight matrix of the compression layer. For bias vectors, To cut the dimension of the state vector, This is the output sequence.

7. The method for detecting abnormalities in the cutting accuracy of a pelletizer as described in claim 6, characterized in that, The dynamic threshold comparison specifically includes: Cutting state vector Input anomaly scorer to calculate Mahalanobis distance ,when Detecting anomalies in real time; in The mean of the historical normal state vector. express Cut the state vector at any time. For transpose operation, Let covariance matrix be the variance matrix. for Dynamic threshold at any given time.

8. The method for detecting abnormalities in the cutting accuracy of a pelletizer as described in claim 7, characterized in that, The dynamic threshold is updated via an adaptive algorithm: Take the nearest Calculate the moving average of the state vector based on hourly normal operating data. and standard deviation The dynamic threshold update formula is: in Based on the basic threshold coefficient, The time decay factor, This represents the number of hours the system has been running.

9. The method for detecting abnormalities in the cutting accuracy of a pelletizer as described in claim 8, characterized in that, The basic threshold coefficient Dynamically adjust based on raw material properties: in , For adjustment coefficients, This represents the hardness value of the current batch of raw materials. This is the viscosity value of the current batch of raw materials. , These are the baseline values ​​for standard raw material parameters.

10. A system for detecting abnormalities in the cutting accuracy of a pelletizer, characterized in that, The system includes: The signal acquisition module collects multi-source signals in real time during the operation of the pelletizer, including vibration signals from the cutting device and three-phase load current signals from the drive motor. The feature extraction module performs time-frequency domain transformation on the vibration signal to extract the impact feature spectrum, and simultaneously performs harmonic decomposition on the three-phase load current signal to generate current distortion factor. The feature fusion module inputs the impact feature spectrum and current distortion factor into a pre-trained deep learning model and generates a cutting state vector through temporal feature fusion. The anomaly detection module determines that the cutting accuracy is abnormal based on the comparison result between the cutting state vector and the dynamic threshold, wherein the dynamic threshold is updated based on historical normal working condition data through an adaptive algorithm. The alarm triggering module triggers an alarm and marks the raw material batch information associated with the abnormal period when an anomaly is detected.

Images