A method for judging the degree of wear of crusher cutters based on BP neural network

Abstract

The present invention discloses a method for judging the degree of wear of crusher tools based on a BP neural network, which relates to the technical field of industrial predictive maintenance. The method comprises the following steps: obtaining original multimodal data and performing preprocessing; performing adversarial training on the preprocessed multimodal data using a generative adversarial network to generate synthetic wear data, and mixing the synthetic wear data with the preprocessed multimodal data to obtain an enhanced training data set; constructing an initial collaborative matrix using the physical layout relationship of the tools, learning historical wear data of multiple tools through a BP neural network, updating weight parameters of the collaborative matrix, obtaining a dynamic collaborative influence matrix, performing tensor splicing of the independent features of each tool with the dynamic collaborative influence matrix, and obtaining a multi-tool fusion feature vector; constructing a multi-task BP neural network model, and training the multi-task BP neural network model using the enhanced training data set.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial predictive maintenance, and in particular to a method for judging the wear degree of a crusher tool based on a BP neural network. Background Art

[0002] Traditional methods for monitoring crusher tool wear rely primarily on single-modal sensor data (such as vibration or temperature) combined with shallow machine learning models. This approach suffers from two significant drawbacks. First, the distribution of tool wear samples over the entire lifecycle in industrial scenarios is extremely uneven, with a disproportionate proportion of normal wear data and a severe lack of moderate to severe wear samples. This makes acquiring effective data particularly difficult under extreme operating conditions, such as high loads and foreign object jamming.

[0003] The synergistic wear effect of multiple cutters has long been unquantified. Due to the close physical proximity of crusher cutters during operation, wear on a single cutter can accelerate wear on adjacent cutters through mechanisms such as stress conduction and load redistribution. Existing technologies generally analyze data from individual cutters independently. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method for judging the degree of wear of crusher tools based on BP neural network to solve the problems of insufficient model generalization caused by scarcity of tool wear samples and difficulty in quantifying the collaborative wear effect of multiple tools.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for determining the degree of wear of a crusher tool based on a BP neural network, which comprises obtaining original multimodal data and performing preprocessing;

[0008] A generative adversarial network is used to perform adversarial training on the preprocessed multimodal data to generate synthetic wear data, which is then mixed with the preprocessed multimodal data to obtain an enhanced training dataset.

[0009] The initial synergy matrix is ​​constructed using the physical layout relationship of the tools. The historical wear data of multiple tools are learned through a BP neural network to update the weight parameters of the synergy matrix and obtain a dynamic synergy influence matrix. The independent features of each tool are tensor-joined with the dynamic synergy influence matrix to obtain a multi-tool fusion feature vector.

[0010] Construct a multi-task BP neural network model and train it using the enhanced training data set;

[0011] The multimodal data collected in real time is input into the trained BP neural network model to obtain the current wear degree judgment result.

[0012] As a preferred solution of the method for judging the degree of wear of crusher tools based on BP neural network described in the present invention, the preprocessing refers to frequency domain denoising of vibration and sound wave signals, normalization of temperature and power data, and standardized size cropping and grayscale conversion of image data.

[0013] As a preferred solution of the method for judging the degree of wear of crusher tools based on BP neural network described in the present invention, the generative adversarial network is a conditional generative adversarial network, which generates synthetic wear data covering different wear levels and working conditions by inputting wear state condition vectors.

[0014] As a preferred solution of the method for judging the wear degree of crusher cutters based on BP neural network of the present invention, wherein: the construction of the synergy matrix refers to initializing weight parameters using the adjacent relationship of the physical layout of the cutters, and the initial weights of adjacent cutters are higher than those of non-adjacent cutters;

[0015] The weight update of the dynamic collaborative influence matrix is ​​carried out by nonlinear mapping learning of the temporal correlation of multi-tool historical wear data through BP neural network.

[0016] As a preferred solution of the method for judging the wear degree of crusher tools based on BP neural network of the present invention, wherein: the BP neural network takes the independent features of the historical wear data of multiple tools as input and the actual wear correlation between the tools as the supervision label;

[0017] The update process of the dynamic collaborative influence matrix uses the mean square error loss function to optimize the matrix weights.

[0018] As a preferred solution of the method for judging the wear degree of crusher tools based on BP neural network of the present invention, wherein: the feature fusion layer performs multi-dimensional tensor splicing on the independent feature vectors of each tool and the flattened vectors of the dynamic synergistic influence matrix;

[0019] The fused feature vector is processed through a fully connected layer for dimensionality reduction, which retains the high-order correlation information of the individual characteristics of the tool and the synergistic influence.

[0020] As a preferred solution of the method for judging the wear degree of crusher cutter based on BP neural network of the present invention, wherein: the multi-task BP neural network model is composed of a first output branch and a second output branch as a core;

[0021] The first output branch outputs the current wear level of each tool through the Softmax classifier;

[0022] The second output branch predicts the wear rate and remaining life in the future time period through the LSTM network.

[0023] As a preferred solution of the method for judging the wear degree of crusher cutters based on BP neural network of the present invention, the shared hidden layer of the multi-task BP neural network model is composed of a multi-layer fully connected network, which is activated and initialized using the ReLU activation function and the Xavier weight initialization method;

[0024] The outputs of the shared hidden layer are input to the Softmax classifier and the LSTM regression network respectively.

[0025] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the method for determining the degree of wear of a crusher tool based on a BP neural network as described in the first aspect of the present invention is implemented.

[0026] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, any step of the method for determining the degree of wear of crusher tools based on a BP neural network as described in the first aspect of the present invention is implemented.

[0027] The beneficial effects of the present invention are as follows: the present invention constructs a multimodal data enhancement mechanism through generative adversarial networks, effectively solving the problems of scarcity and uneven distribution of tool wear samples in industrial scenarios, generating data covering the entire wear cycle and extreme working conditions, and significantly expanding the sample diversity of model training; constructing a dynamic collaborative influence matrix based on the physical layout of the tool, quantifying the stress conduction and load redistribution effects between multiple tools through the BP neural network, breaking through the limitations of traditional single-tool independent analysis, and realizing accurate modeling of collaborative wear of tool groups; innovatively designing a classification-regression dual-branch multi-task model, synchronously outputting the current wear level and future life prediction, forming a closed-loop decision-making system from real-time monitoring to preventive maintenance, greatly improving the accuracy of abnormality detection and maintenance efficiency, and reducing equipment operation and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0029] Figure 1 This is a flowchart of multimodal data processing and enhancement in Example 1.

[0030] Figure 2Schematic diagram of the dynamic collaborative matrix update mechanism in Example 1.

[0031] Figure 3 This is a diagram of the multi-task neural network architecture in Example 1.

[0032] Figure 4 This is a flow chart of real-time monitoring and feedback in Example 1. DETAILED DESCRIPTION

[0033] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0034] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0035] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0036] Example 1, reference Figures 1 to 4 , which is the first embodiment of the present invention, provides a method for judging the degree of wear of crusher cutters based on BP neural network, comprising the following steps:

[0037] S1. Obtain original multimodal data and perform preprocessing.

[0038] Specifically, multi-source sensors installed on the crusher collect real-time data on the operating status of multiple cutters, generating raw multimodal data including vibration signals, temperature data, power consumption, acoustic signals, and cutter surface images. Specifically, sensors were deployed on a crusher with eight cutters to collect multi-dimensional data during operation. This raw multimodal data was then preprocessed to ensure data quality.

[0039] Furthermore, each tool is equipped with a three-axis acceleration sensor (model ADXL345), with a sampling frequency set to 1000 Hz and a measuring range of ±16g, to collect vibration acceleration data of the tool when crushing the material.

[0040] Each tool is equipped with a thermocouple (K-type thermocouple) with a sampling frequency of 1 Hz and a measurement range of 0-500°C to collect tool surface temperature data.

[0041] Install a power meter (Fluke 1732) on the crusher motor with a sampling frequency of 1 Hz and a measurement range of 0-50 kW to collect power data when the motor is running.

[0042] Install a high-sensitivity microphone (Brüel&Kjær 4189) near the crusher cutter head with a sampling frequency of 1000 Hz to collect the sound wave signals when the cutter is working.

[0043] Install an industrial high-definition camera (model Basler acA1300-30gm) inside the crusher with a resolution of 1280×960 and a shooting frequency of once per minute to capture images of the tool surface.

[0044] Data collection is centrally controlled by a PLC (e.g., a Siemens S7-1200), which reads data from each sensor every second and stores it in a local database (formatted in SQLite). Collection lasts for 8 hours per shift, generating approximately 1GB of data per day.

[0045] Wavelet transform (using the Daubechies 4 wavelet basis) is used on vibration and acoustic signals to remove high-frequency noise while retaining the main frequency components. This removes high-frequency noise from vibration and acoustic signals while retaining the main wear characteristic frequency bands (such as tool resonance frequency).

[0046] Temperature, power, and vibration amplitude are normalized to the range [0, 1] using the minimum-maximum normalization method. This unifies the dimensional differences in temperature, power, and vibration amplitude, accelerating convergence in subsequent training. The tool edge ROI (maximum 200×200 pixels) is extracted from the image data using the OpenCV library. After grayscale conversion, it is resized to a standardized 64×64 pixel matrix. Multi-source data is aligned by timestamps, with an error of less than ±10ms. This allows for focused analysis of the tool edge wear region, eliminating background interference and improving the efficiency of subsequent visual feature extraction.

[0047] It should be noted that this step achieves comprehensive multimodal data acquisition through multi-source sensors, covering both the physical characteristics of the tool's operating state (such as vibration and temperature) and environmental influences (such as acoustic waves and images), providing multi-dimensional data support for wear analysis. Preprocessing ensures data cleanliness and consistency, providing high-quality input for subsequent data enhancement and model training. The sampling frequency is selected based on the dynamic characteristics of tool wear signals (vibration and acoustic waves require high-frequency capture, while temperature and power change more slowly). Image resolution and ROI size balance computational efficiency with feature preservation. The raw multimodal dataset generated approximately 7,200 vibration and acoustic wave samples, 480 temperature and power samples, and 480 images per tool per day, providing a sufficient data foundation for subsequent steps. The grayscale value mapping algorithm utilizes FFT spectrum amplitude normalization.

[0048] S2. Use a generative adversarial network to perform adversarial training on the preprocessed multimodal data to generate synthetic wear data, and mix the synthetic wear data with the preprocessed multimodal data to obtain an enhanced training dataset.

[0049] Specifically, a generative adversarial network (GAN) is used to perform adversarial training on the acquired multimodal data to generate synthetic wear data. This synthetic wear data is then mixed with real data (i.e., preprocessed multimodal data) to obtain an enhanced training dataset. Specifically, a conditional generative adversarial network (cGAN) is used to generate simulated tool wear data covering light, moderate, and heavy wear, as well as extreme operating conditions (such as high loads and foreign object jamming). This compensates for the lack of samples or uneven distribution in the real data, ultimately forming an enhanced dataset suitable for model training.

[0050] Furthermore, GAN consists of a generator and a discriminator.

[0051] The generator is structured as a five-layer fully connected neural network, consisting of an input layer (100-dimensional random noise + a five-dimensional conditional vector), hidden layer 1 (256 neurons, ReLU activation), hidden layer 2 (512 neurons, ReLU activation), hidden layer 3 (256 neurons, ReLU activation), and an output layer (matching the multimodal data dimensions, such as 64-dimensional vibration and 1-dimensional temperature). By inputting random noise and a wear conditional vector (a five-dimensional one-hot encoding), it generates multimodal synthetic data (vibration, temperature, power, sound waves, and images) covering conditions from "normal" to "failure," overcoming the limitation of insufficient samples of extreme operating conditions in real data.

[0052] A 100-dimensional random noise vector (following the normal distribution N(0,1)) and a 5-dimensional conditional vector (one-hot encoding, representing the five states of "normal", "light wear", "moderate wear", "heavy wear" and "failure") are used as input to the generator.

[0053] The simulated multimodal data, including vibration signals (64-dimensional spectral features), temperature (1-dimensional), power (1-dimensional), sound waves (64-dimensional spectral features), and images (64×64 pixel matrices), are used as the output of the generator.

[0054] The structure of the discriminator is a three-layer convolutional neural network, with the layers being the input layer (receiving multimodal data), convolutional layer 1 (with 32 3×3 filters, stride set to 1, activated by LeakyReLU), convolutional layer 2 (with 64 3×3 filters, stride set to 2, also activated by LeakyReLU), and output layer (1D, Sigmoid activation).

[0055] The input of the discriminator is the real data and the synthetic data output by the generator.

[0056] The discriminator outputs a probability value between 0 and 1, indicating the likelihood that the data is true. Convolutional layers extract local correlation features from multimodal data (such as temporal patterns in vibration spectra or wear patterns in image textures), accurately identifying distributional differences between generated and real data, and driving the generated data to approximate real-world physical laws.

[0057] The training process is as follows: 500 sets of real data are selected as the initial training set, with approximately 100 sets for each wear state. The Adam optimizer is used with a learning rate of 0.0002, momentum parameters β1 set to 0.5, and β2 set to 0.999. A batch size of 32 is used for 1000 epochs. Each epoch first updates the discriminator five times to enable it to distinguish between real and synthetic data. The generator is then updated once to ensure it produces more realistic data. Every 100 epochs, the generator model is saved and the quality of the generated data is evaluated. This balance between gradient update speed and stability is maintained to accelerate convergence and reduce mode collapse in the generated data (e.g., logical inconsistencies between vibration signals and temperature data).

[0058] Using the trained generator, different conditional vectors (such as "heavy wear + high load") were input to generate 1,000 sets of synthetic data, each containing multimodal features including vibration, temperature, power, acoustic waves, and images. The generated data volume was five times the real data, totaling 2,500 sets (500 real sets + 2,000 synthetic sets). By prioritizing the discriminator for five iterations per round, the discriminator's discriminative power was improved, preventing the generator from prematurely falling into local optima and ensuring the diversity of the generated data. The KL divergence (Kullback-Leibler divergence) was used to calculate the distributional similarity between the generated data and the real data, with a target value of less than 0.05. Statistical analysis (mean and variance) of the generated data was performed to ensure consistency with the real data. The cross-modal correlation between the vibration spectrum (64 dimensions), acoustic signal (64 dimensions), and image matrix (64×64 pixels) in the generated data was verified to avoid model training bias caused by generating isolated modal data.

[0059] 2500 sets of augmented data (consisting of 500 sets of real data + 2000 sets of combined numbers) are randomly shuffled to form an augmented training data set and stored as an HDF5 format file.

[0060] It should be noted that the generation of multimodal data using conditional GANs addresses the scarcity of tool wear samples in industrial scenarios. The application of the Wasserstein distance improves the stability and authenticity of the generated data, while the introduction of conditional vectors ensures the relevance of the generated data to specific wear states. The generation of an enhanced dataset not only expands the training sample size but also covers extreme operating conditions, improving the generalization of subsequent models. The number of network layers and neurons in the generator and discriminator is set based on the complexity of the multimodal data, ensuring a balance between computational efficiency and generation quality. Training parameters (such as a learning rate of 0.0002 and 1000 rounds) are based on empirical values ​​from GANs used in industrial data generation. A KL divergence threshold of 0.05 ensures high similarity in data distribution. The scale of generating 1000 data sets is a fivefold increase from the daily data volume (approximately 480 sets) for eight crusher tools, meeting training requirements. An enhanced training dataset consisting of 2,500 sets of multimodal data, each containing vibration (64 dimensions), temperature (1 dimension), power (1 dimension), sound waves (64 dimensions), and images (64×64 pixels), provides rich data support for subsequent collaborative feature extraction and model training. The collaborative optimization design of the generator and discriminator, enhanced stability of the adversarial training strategy, and cross-modal data consistency verification ensure the physical plausibility and statistical authenticity of the synthesized data.

[0061] S3, using the tool physical layout relationship to build the initial collaborative matrix and obtain the multi-tool fusion feature vector,

[0062] Specifically, an initial synergy matrix is ​​constructed using the physical layout of the tools. A BP neural network is then used to learn from historical wear data for multiple tools, updating the weight parameters of the synergy matrix to obtain a dynamic synergy influence matrix. Subsequently, a feature fusion layer is used to tensor-concatenate the independent features of each tool with the dynamic synergy influence matrix to obtain a multi-tool fusion feature vector. For example, on a crusher with eight tools, the collaborative wear characteristics between the tools were extracted based on an enhanced training dataset, reflecting the interaction between individual wear and group wear.

[0063] Furthermore, in the initial coordination matrix construction step, the matrix is ​​defined as an 8×8 initial coordination matrix M (the format is preferably a symmetric matrix), the elements Represents the influence weight of the wear of the i-th tool on the j-th tool.

[0064] The initialization rule is adjacent tools (physical distance less than 10cm) Set to 0.7. Set to 0.3. Diagonal elements are fixed at 1.0 (indicating the influence of the cutter itself). The data source is the physical layout parameters of the crusher cutterhead. The collaborative weights are initialized based on the actual cutter installation positions (adjacent distances less than 10 cm), assigning higher influence coefficients to adjacent cutters (0.7 for non-adjacent cutters, 0.3 for adjacent cutters). This avoids deviations from physical laws caused by random initialization and improves the rationality of the initial matrix.

[0065] The BP neural network is used to update the coordination matrix. The structure of the BP neural network includes an input layer (8×130-dimensional historical data), a hidden layer (128→64 neurons, activated by the ReLU activation function), and an output layer (64-dimensional matrix flattened). The training process uses 1000 sets of historical data and annotates the true correlation coefficient. , the loss function is mean square error (MSE), and the optimization parameters are Adam (learning rate 0.001, batch size 32, 500 rounds).

[0066] The feature extraction and fusion process begins with independent feature extraction: for each tool, extracting the vibration spectrum (64-dimensional FFT), temperature (1-dimensional), power (1-dimensional), and acoustic spectrum (64-dimensional FFT), for a total of 130 dimensions. Feature fusion is then performed, concatenating the independent features of the eight tools (8 × 130 dimensions) with the dynamic matrix M' (64-dimensional flattened) to form a 1104-dimensional input. This is then reduced to 256 dimensions using a fully connected layer (ReLU activation, Xavier initialization).

[0067] It should be noted that quantitative modeling of inter-tool wear effects is achieved through the dynamic updating of the initial synergy matrix and the BP neural network. The initial matrix provides prior knowledge based on the physical layout, while the BP network learns complex dependencies from historical data. The resulting dynamic synergy influence matrix, M', reflects the true synergy between tools. The feature fusion layer combines individual features with group effects to form a comprehensive multi-tool fusion feature vector, providing high-dimensional input for subsequent wear prediction.

[0068] Initial matrix weights (0.7 and 0.3) were chosen based on empirical values ​​of the tool's physical distance to ensure reasonableness. The number of BP network layers and neurons (128 and 64) balanced computational complexity and fitting performance. A learning rate of 0.001 and 500 training rounds were used, reflecting conventional settings for industrial data processing. The fusion layer output was 256-dimensional, effectively reducing the high-dimensional input (1104-dimensional) while preserving key feature information.

[0069] The resulting multi-tool fusion feature vector has a dimension of 256, with each set of data corresponding to the collaborative wear state of eight tools. This data provides input for subsequent multi-task BP neural network training. The dynamic collaborative influence matrix M' (8×8) can be used independently to analyze the interaction patterns between tools.

[0070] S4. Construct a multi-task BP neural network model and train it using the enhanced training data set.

[0071] Specifically, a multi-task BP neural network model was constructed and trained using an enhanced training dataset. The model consists of two output branches: the first output branch outputs the current wear level of each tool through a Softmax classifier, accurately identifying the current wear level of the tool (normal to failure) through discrete label supervised learning; the second output branch predicts the wear rate and remaining life in the future time period through an LSTM network. The LSTM network captures the temporal dependency of the wear rate and remaining life (with a 10-minute time window), predicting the wear trend within the next 8 hours with a remaining life error of less than 5 hours, providing a quantitative basis for preventive maintenance. Specifically, based on the multi-tool fusion feature vector, a multi-task neural network was constructed and trained to achieve real-time classification and trend prediction of the wear status of 8 tools.

[0072] Furthermore, the network structure includes a 256-dimensional input layer to receive fused feature vectors (each set of data corresponds to the collaborative wear characteristics of 8 tools); three fully connected layers serve as shared hidden layers: (Layer 1: fully connected layer, 128 neurons, ReLU activation function, weight initialization uses the Xavier method to balance the distribution of activation values ​​​​of each layer, accelerate the convergence speed (reducing the number of training rounds by 20%), and alleviate the gradient vanishing problem to ensure the stability of the deep network. Layer 2: fully connected layer, 64 neurons, ReLU activation function. Layer 3: fully connected layer, 32 neurons, ReLU activation function.) Through three layers of ReLU-activated shared hidden layers, the common laws in the fusion characteristics of multiple tools (such as wear thermal coupling effect and group load transfer mode) are abstracted step by step, thereby improving the ability to represent complex correlation relationships.

[0073] Fusion features were extracted from the augmented training dataset (2,500 sets), each set consisting of a 256-dimensional fused feature vector and a corresponding label. The wear level of each tool was annotated using historical maintenance records, with wear levels categorized into 5 levels (1 for normal wear and 5 for severe wear). The wear rate (based on the rate of change of vibration amplitude) and remaining life (based on a threshold time) were calculated from the historical data to determine the regression label. The data was divided into 80% (2,000 sets) as the training set and 20% (500 sets) as the validation set.

[0074] The optimization parameters are as follows: the Adam optimizer is used, the learning rate is set to 0.001, and the momentum parameters are set to β1=0.9 and β2=0.999.

[0075] The batch size was 32 samples, and the training epochs were 500. Early stopping was required to terminate training if the validation set loss did not decrease after 10 consecutive epochs. Training was dynamically terminated based on the validation set loss to prevent overfitting. 2000 training data sets were fed into the network, and forward propagation was used to calculate the classification and regression outputs. Backward propagation was used to calculate the gradients and update the weights of the shared and branch layers. Classification accuracy and regression error were evaluated on the validation set every 50 epochs, and the optimal model was saved. The classification accuracy target on the validation set was greater than 90%, and a confusion matrix was used to display the predicted distribution of each class. The wear rate prediction error was less than 0.1% / hour, and the remaining life prediction error was less than 5 hours.

[0076] It should be noted that this step achieves the dual goals of tool wear classification and trend prediction through a multi-task BP neural network. The shared hidden layer extracts common features, the classification branch provides discrete judgments about the current state, and the LSTM branch captures temporal dependencies to predict future trends. The multi-task loss function balances the training of the classification and regression tasks, ensuring balanced model performance across both tasks.

[0077] The number of hidden layers (3) and neurons (128, 64, and 32) were set based on feature complexity, and the dimensionality was gradually reduced to avoid overfitting.

[0078] The LSTM time window (10 minutes) is determined according to the dynamic change cycle of tool wear, and the 32 hidden units balance computational efficiency and prediction accuracy.

[0079] The number of training rounds and early stopping strategy refer to the experience of industrial neural network training to ensure model convergence.

[0080] The trained multi-task BP neural network model can input a 256-dimensional fused feature vector and output the current wear level (five-level classification), future wear rate, and remaining life (regression value) of eight tools, supporting subsequent real-time judgment. The modular design of the shared hidden layer supports subsequent online fine-tuning (such as a learning rate of 0.0001), adapting to new tool materials or working conditions by simply updating branch layer parameters.

[0081] S5. Input the multimodal data collected in real time into the trained BP neural network model to obtain the current wear degree judgment result.

[0082] Specifically, real-time multimodal data collected is fed into a trained BP neural network model to determine the current wear level. When tool wear exceeds a preset threshold or the remaining life is insufficient, an alarm is triggered and a maintenance work order is generated. Specifically, the trained multi-task BP neural network is deployed on the crusher's edge computing device, using real-time sensor data for online predictions. An automated feedback mechanism optimizes model performance, enabling real-time monitoring of the wear status of the eight tools and enabling maintenance decisions.

[0083] Furthermore, the trained BP neural network model (in PyTorch format) was converted to the ONNX format. ONNX optimization tools (such as onnx-simplifier) ​​were used to simplify the model structure and reduce computational overhead. This model format conversion and structural optimization (such as operator fusion) reduced inference computation by 40%, enabling 10 real-time predictions per second (less than 100ms) on an edge device (Jetson Nano), meeting the high concurrency requirements of industrial scenarios. The BP neural network model is approximately 10MB in size, with an inference time target of less than 100ms. The model was deployed on an industrial edge computing module (e.g., the NVIDIA Jetson Nano). It connected to a PLC (Siemens S7-1200) via the Modbus protocol to receive real-time sensor data.

[0084] One set of multimodal data (vibration 1000Hz, temperature 1Hz, etc.) is collected every second, and features are extracted using a sliding window (10 seconds). Time series segments of vibration and acoustic signals (1000Hz x 10 seconds = 10,000 points) are dynamically captured to extract transient impact features (such as peak value and energy entropy). This enhances the model's response sensitivity to sudden anomalies (such as foreign object jamming), and limits false alarm delays to less than 2 seconds.

[0085] A 256-dimensional feature vector is fed into the model every second, and the forward propagation outputs two sets of results: Branch 1: The wear level of the eight tools (5 levels: normal, light, moderate, severe, failure). Branch 2: The wear rate (% / hour) and remaining life (hours) of the eight tools.

[0086] The HMI interface (e.g., Siemens TP700 Comfort) displays in real time the current wear level (e.g., "Tool 1: Slight, Tool 2: Moderate") and trend predictions (e.g., "Tool 1: Rate 0.5% / hour, Life 50 hours"). The data refresh rate is 1Hz.

[0087] Set the wear threshold to predict any tool as "critical" or "failed". The life threshold is when the remaining life is less than 8 hours and the wear rate is greater than 2% / hour.

[0088] When any threshold condition is met, the edge device automatically triggers an audible and visual alarm (model: Patlite LR6) via the GPIO interface, emitting an audible and visual signal (85 decibels, flashing red light). Simultaneously, a maintenance work order is automatically generated as a JSON file (containing the tool number, wear status, and recommended maintenance time) and pushed to the Manufacturing Execution System (MES) via Ethernet. This dual judgment, combining wear level (Softmax output) and life prediction (LSTM output), prevents over-maintenance and under-reporting.

[0089] Automatically compares predictions with actual conditions (real data obtained through regular maintenance records or sensor anomaly detection), recording deviations. Accumulated deviations trigger a lightweight update of model parameters (learning rate 0.0001), enabling the model to continuously adapt to tool performance degradation and new operating conditions, keeping long-term accuracy fluctuations within ±3%.

[0090] Using an online gradient descent algorithm, model fine-tuning is triggered when the cumulative deviation samples reach 32 sets per day, with a learning rate of 0.0001, to update the weights of the shared hidden layer and branch layers. Model prediction accuracy (classification accuracy greater than 90%, lifetime error less than 5 hours) and response time (less than 100ms) are recorded hourly. If the accuracy falls below 85% for three consecutive days, an offline retraining request is automatically generated and pushed to the cloud server.

[0091] It should be noted that this step deploys the trained multi-task BP neural network to edge devices, enabling real-time wear assessment and prediction for industrial applications. Automated alarms and maintenance work order generation provide closed-loop support from prediction to decision-making, while an online feedback mechanism automatically optimizes the model through machine learning, ensuring long-term adaptability. HMI interface and MES integration enhance practicality. The model's inference time target (less than 100ms) and refresh rate (1Hz) are based on industrial real-time requirements to ensure no impact on crusher operation. Threshold settings (severity level, 8-hour lifespan, 2% rate) are based on tool maintenance experience, balancing warning timeliness and false alarm rate. Online learning parameters (learning rate 0.0001, 32 samples) are lightweight updates to avoid impacting real-time performance. The hardware choice (Jetson Nano) balances cost and computing power, making it suitable for edge deployment.

[0092] Get real-time updates on tool wear assessment results (5-level classification for 8 tools), trend predictions (wear rate and remaining life), and alarm signals and work orders. Automated feedback maintains high accuracy and provides intelligent support for crusher maintenance.

[0093] This embodiment also provides a computer device, which is suitable for the method for judging the degree of wear of crusher tools based on BP neural network, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the method for judging the degree of wear of crusher tools based on BP neural network proposed in the above embodiment.

[0094] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0095] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for determining the degree of wear of crusher cutters based on a BP neural network as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0096] In summary, the present invention constructs a multimodal data enhancement mechanism through generative adversarial networks, effectively solving the problems of scarcity and uneven distribution of tool wear samples in industrial scenarios, generating data covering the entire wear cycle and extreme working conditions, and significantly expanding the sample diversity of model training; constructing a dynamic collaborative influence matrix based on the physical layout of the tool, and quantifying the stress conduction and load redistribution effects between multiple tools through the BP neural network, breaking through the limitations of traditional single-tool independent analysis, and realizing accurate modeling of collaborative wear of tool groups; innovatively designing a classification-regression dual-branch multi-task model, synchronously outputting the current wear level and future life prediction, forming a closed-loop decision-making system from real-time monitoring to preventive maintenance, greatly improving the accuracy of abnormality detection and maintenance efficiency, and reducing equipment operation and maintenance costs.

[0097] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for judging the degree of wear of crusher cutters based on BP neural network, characterized by: include, Obtain raw multimodal data and perform preprocessing; A generative adversarial network is used to perform adversarial training on the preprocessed multimodal data to generate synthetic wear data, which is then mixed with the preprocessed multimodal data to obtain an enhanced training dataset. The initial synergy matrix is ​​constructed using the physical layout relationship of the tools. The historical wear data of multiple tools is learned through a BP neural network to update the weight parameters of the initial synergy matrix and obtain a dynamic synergy influence matrix. The independent features of each tool are tensor-joined with the dynamic synergy influence matrix to obtain a multi-tool fusion feature vector. Construct a multi-task BP neural network model and train it using the enhanced training data set; Input the multimodal data collected in real time into the trained BP neural network model to obtain the current wear degree judgment result; The construction of the initial synergy matrix refers to initializing weight parameters using the adjacent relationship of the tool physical layout, where the initial weight of adjacent tools is higher than that of non-adjacent tools; The weight update of dynamic collaborative influence matrix is ​​carried out by nonlinear mapping learning of the temporal correlation of multi-tool historical wear data through BP neural network; The BP neural network takes the independent features of historical wear data of multiple tools as input and the actual wear correlation between tools as supervision labels.

2. The method for judging the degree of wear of crusher cutters based on BP neural network according to claim 1, characterized in that: The preprocessing refers to performing frequency domain denoising on vibration and sound wave signals, normalizing temperature and power data, and performing standardized size cropping and grayscale conversion on image data.

3. The method for judging the degree of wear of crusher cutters based on BP neural network according to claim 2, characterized in that: The generative adversarial network is a conditional generative adversarial network that generates synthetic wear data covering different wear levels and working conditions by inputting a wear state condition vector.

4. The method for judging the degree of wear of crusher cutters based on BP neural network according to claim 3, characterized in that: The update process of the dynamic collaborative influence matrix uses the mean square error loss function to optimize the matrix weights.

5. The method for judging the degree of wear of crusher cutters based on BP neural network according to claim 4, characterized in that: The feature fusion layer performs multi-dimensional tensor splicing on the independent feature vectors of each tool and the flattened vectors of the dynamic synergistic influence matrix; The fused feature vector is processed through a fully connected layer for dimensionality reduction, which retains the high-order correlation information of the individual characteristics of the tool and the synergistic influence.

6. The method for judging the degree of wear of crusher cutters based on BP neural network according to claim 5, characterized in that: The multi-task BP neural network model is composed of a first output branch and a second output branch as a core; The first output branch outputs the current wear level of each tool through the Softmax classifier; The second output branch predicts the wear rate and remaining life in the future time period through the LSTM network.

7. The method for judging the degree of wear of crusher cutters based on BP neural network according to claim 6, characterized in that: The shared hidden layer of the multi-task BP neural network model is composed of a multi-layer fully connected network, which is activated and initialized using the ReLU activation function and the Xavier weight initialization method; The outputs of the shared hidden layer are input to the Softmax classifier and the LSTM regression network respectively.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for judging the degree of wear of crusher cutters based on a BP neural network are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for judging the degree of wear of crusher cutters based on a BP neural network according to any one of claims 1 to 7 are implemented.

Images