Electric drill torque dynamic adjusting method and device, electronic equipment, medium and program product

Abstract

The invention belongs to the technical field of electric drill torque dynamic adjustment, and discloses an electric drill torque dynamic adjustment method and device, electronic equipment, a medium and a program product, and the method comprises the steps: obtaining motor operation data and user operation behavior data, and recognizing the type and state of a material through a first deep learning model, the second deep learning model judges the user intention, and the third deep learning model analyzes the drill bit state. Based on the three types of recognition results, a torque control strategy is generated by combining an expert rule base and dynamic weight correction, and dynamic adjustment of the output torque of the electric drill is achieved. The method can adapt to multi-material working conditions and operation scenes, gives consideration to real-time response and safety control, improves the drilling efficiency, and is suitable for electric drill intelligent operation scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of dynamic adjustment of electric drill torque, and in particular to a method, device, electronic equipment, medium and program product for dynamic adjustment of electric drill torque. Background Art

[0002] The torque adjustment of traditional electric drills mainly relies on manual settings or simple rule control (such as current thresholds), which makes it difficult to cope with the multiple uncertainties that arise in complex environments. For example, unknown materials, complex structures (such as multi-layer composite panels, hollow bricks, and old walls), drill bit wear, eccentricity and other non-material factors causing anomalies, different user operation methods, and difficult to judge intentions (whether it is testing, pushing, or misoperation), etc. These factors make traditional torque adjustment prone to misjudgment, leading to problems such as slipping, drill jamming, and accidental damage. Especially in high-risk scenarios such as home decoration and old house renovation, the user experience and safety are poor. Summary of the Invention

[0003] The purpose of the present invention is to address the above-mentioned problems and provide a method, device, electronic equipment, medium and program product for dynamic adjustment of electric drill torque.

[0004] To achieve one of the above-mentioned objectives of the invention, one embodiment of the present invention provides a method for dynamically adjusting the torque of an electric drill, comprising: obtaining motor operation data of the electric drill during operation, including at least time series data of current, voltage, control signal, speed, and motor load response; obtaining a first recognition result based on the motor operation data through a first deep learning model, the first recognition result including identifying the type of drilled material and its status; obtaining user operation behavior data on the electric drill, the operation behavior data including at least the pressing duration, frequency, and rhythm change of the electric drill trigger signal, obtaining a second recognition result based on the operation behavior data through a second deep learning model, the second recognition result identifying the user's current usage intention, including tentative operation, continuous propulsion operation, or impact operation; obtaining a third recognition result based on the motor operation data through a third deep learning model, the third recognition result including identifying the drill bit status; obtaining a control strategy for dynamically adjusting the torque based on the first recognition result, the second recognition result, and the third recognition result, and dynamically adjusting the output torque of the electric drill based on the control strategy.

[0005] To achieve one of the aforementioned objectives of the invention, one embodiment of the present invention provides a device for dynamically adjusting torque of an electric drill, comprising: a first acquisition module for acquiring motor operation data of the electric drill during operation, including at least time series data of current, voltage, control signal, speed, and motor load response; a first identification module for obtaining a first recognition result using a first deep learning model based on the motor operation data, wherein the first recognition result includes identifying the type and status of the drilled material; a second identification module for obtaining user operation behavior data on the electric drill, wherein the operation behavior data includes at least the pressing duration, frequency, and rhythm changes of the electric drill trigger signal, and obtaining a second recognition result using a second deep learning model based on the operation behavior data, wherein the second recognition result identifies the user's current usage intention, including a tentative operation, a continuous push operation, or an impact operation; a third identification module for obtaining a third recognition result using a third deep learning model based on the motor operation data, wherein the third recognition result includes identifying the drill bit status; and an adjustment module for obtaining a control strategy for dynamically adjusting torque based on the first, second, and third recognition results, and dynamically adjusting the output torque of the electric drill based on the control strategy.

[0006] To achieve one of the above-mentioned objects of the invention, an embodiment of the present invention provides a storage medium storing program instructions, which, when executed, implements the electric drill torque dynamic adjustment method as described in any one of the above items.

[0007] To achieve one of the above-mentioned objects of the invention, an embodiment of the present invention provides an electronic device, including a processor and a memory, wherein the memory stores program instructions, and the processor runs the program instructions to implement the electric drill torque dynamic adjustment method as described in any one of the above items.

[0008] Through the solution provided by the embodiment of the present invention, three deep learning models are used to respectively identify the material type and state, user operation intention, and drill bit state. Combined with the expert rule library and dynamic weight correction, accurate identification and dynamic torque adjustment of various materials such as wood and metal, various operation intentions such as probing and impact, and various states of the drill bit such as normal and blunt are achieved, thereby improving the adaptability and efficiency of drilling operations and enhancing user experience and safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is a flow chart of the method for dynamically adjusting the torque of an electric drill according to the present invention.

[0010] Figure 2 This is a schematic diagram of the first deep learning model architecture of the present invention.

[0011] Figure 3 It is a structural schematic diagram of the electric drill torque dynamic adjustment device of the present invention.

[0012] Figure 4 It is a structural schematic diagram of the electronic device of the present invention. DETAILED DESCRIPTION

[0013] The present invention will be described in detail below with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any structural, methodological, or functional changes made by those skilled in the art based on these embodiments are all within the scope of protection of the present invention.

[0014] The scope of the embodiments herein includes the entire scope of the claims, and all available equivalents of the claims. Herein, the terms "first", "second", etc. are only used to distinguish one element from another, without requiring or implying any actual relationship or order between these elements. The terms "include", "comprises", or any other variants thereof are intended to cover non-exclusive inclusions and do not exclude the presence of other identical elements in the structure, device or equipment including the elements. The various embodiments are described in a progressive manner herein, and each embodiment focuses on the differences from other embodiments, and the same similar parts between the various embodiments can be referred to each other.

[0015] An embodiment of the present invention provides a method for dynamically adjusting the torque of an electric drill. Figure 1 As shown, the method includes the following steps: Step 101: obtaining motor operation data of the electric drill during operation, including at least time series data of current, voltage, control signal, speed and motor load response.

[0016] The motor operation data during the operation of the electric drill is the basis for realizing dynamic torque regulation. In some embodiments, the acquisition scope mainly covers the motor electrical parameters and mechanical response characteristics; among which, the electrical parameters include current and voltage data, and PWM control signals; the current and voltage are the real-time current (unit: A) and voltage (unit: V) signals when the motor is working. The current is collected by a Hall current sensor connected in series in the motor circuit, and the voltage is collected by a voltage divider resistor network and a differential amplifier combination circuit; the PWM control signal is the pulse width modulation signal output by the drive circuit. The high-level duration of the PWM waveform can be directly captured through the timer module of the microcontroller, and the frequency (usually 20-50kHz) and duty cycle change rate can be recorded synchronously.

[0017] Mechanical response characteristics primarily include speed data and time series data of the motor's load response. Speed data, i.e., the real-time speed of the motor rotor (unit: rpm), can be calculated by counting pulses using an incremental encoder (e.g., 1024 lines) mounted on the motor shaft, or by inferring the speed by detecting the frequency of the armature voltage using the back-electromotive force method. This is not a limitation in this embodiment. Time series data of the motor's load response primarily includes the peak torque rate of change (unit: N·m / s), frequency perturbation (unit: Hz), and the slope of the transient current rise at the moment of a sudden load change. Torque can be measured using a strain-gauge torque sensor with a sampling frequency of 1 kHz. The frequency domain characteristics of the current signal can be analyzed using a fast Fourier transform (FFT), which is not a limitation in this embodiment.

[0018] Step 102: Based on the motor operation data, a first recognition result is obtained through a first deep learning model, where the first recognition result includes identifying the type of drilled material and its status.

[0019] In this embodiment, Figure 2 As shown, the first deep learning model adopts a lightweight time series deep learning architecture, which is optimized for the time series characteristics of electric drill motor data. For example, it mainly includes: feature extraction layer: one-dimensional convolutional neural network (1D-CNN) convolution kernel (size 3-7) captures local features of signals such as current and torque.

[0020] Temporal modeling layer: Lightweight Transformer encoder (layers 2-4) or LSTM unit (hidden layer dimension 64) to learn long-range dependencies.

[0021] Classification prediction layer: Fully connected layer + Softmax output material type and state label. It is understandable that the convolution kernel size, number of encoder layers, and hidden layer dimension can be set according to actual needs and are not limited in this embodiment.

[0022] In some embodiments, the process of identifying material type and state includes: 1. Data input and preprocessing.

[0023] Slice the motor operation data into time windows (for example, 500ms) to form a multidimensional time series, such as [current series, voltage series, speed series, torque series, PWM series]. Perform Z-Score normalization on the data in each dimension using the formula: x'=(x-μ) / σ, where μ is the mean of the training set and σ is the standard deviation.

[0024] 2. Model training and feature learning.

[0025] Training data: Drilling data of various materials such as wood, metal, and hollow bricks are collected, with each type of sample having no less than 2,000 time windows.

[0026] Feature learning example: Metal material: The current signal shows periodic spikes of 10-15A, and the torque fluctuation frequency is 100Hz.

[0027] Wood: The current is stable (5-8A), the torque change rate is ≤1N・m / s, and the speed fluctuation is small.

[0028] Hollow bricks: The current drops by 30% when the material changes suddenly, and the speed recovers instantly.

[0029] The specific process of model training can be based on existing technologies and will not be described in detail in this embodiment.

[0030] 3. Real-time classification process.

[0031] Sliding window sampling: For example, a 500ms window is slid with a step size of 100ms, and new data is continuously input.

[0032] Feature extraction: 1D-CNN extracts local features such as the current rising edge slope and torque peak interval.

[0033] Temporal reasoning: Transformers learn temporal patterns in how material resistance changes (e.g., the periodic resistance of metals).

[0034] Label output: The Softmax layer outputs the material type and state probability distribution, and takes the maximum probability label.

[0035] Specifically, as an example, sliding window sampling specifically includes: (1) Sampling strategy design: window length and step size: For example, the fixed window length T = 500ms corresponds to the number of motor data sampling points N = T × f_s, where N = 500 points when the sampling frequency f_s = 1kHz; the sliding step size is set to 100ms, that is, 400 new data points are updated every 100ms, and 100 historical data points are retained to form overlapping windows; it can be simplified to the input sequence of the kth window as follows: , where s=100 is the step size and T=500 is the window size.

[0036] (2) Multi-dimensional data fusion: Input tensor structure: As an example, W=500 is the time dimension, and D=5 is the feature dimension (current, voltage, speed, torque, PWM). It can be understood that the W value corresponds to the window length and step size, and the value of DD can be based on actual conditions and is not limited in this embodiment.

[0037] Specifically, as an example, the feature extraction process can include: (1) convolution kernel design: using a 3-layer 1D-CNN with convolution kernel sizes of 7, 5, and 3, and the number of channels of 16, 32, and 64, respectively; lightweight improvement: using depthwise separable convolution, the computational complexity is reduced to 1 / [D×K] of the standard convolution, where K=7 is the kernel size.

[0038] (2) Feature extraction formula can be simplified as: ;in, is the feature map of the lth layer, is the convolution kernel, represents the convolution operation, is the ReLU activation function.

[0039] Specifically, as an example, the process of temporal reasoning can include: (1) Encoder structure: using a 2-layer Transformer encoder, each layer contains multi-head attention and feedforward network; for example, the number of heads h = 4, the embedding dimension d_model = 64, and the feedforward neural network (FFN) dimension d_ff = 128; Lightweight improvement: remove the position encoding and use the convolution layer to retain the temporal information.

[0040] (2) Self-attention mechanism can be simplified into the formula: ; Among them, Q (Query, query vector), , N is the sequence length (e.g. the number of motor data points in a 500ms window), is the vector dimension (here , indicating that the query intent is encoded using a 16-dimensional vector); for example, at a certain moment, Q focuses on "whether the current change matches the characteristics of the metal material" and uses this "search intent" to match other information.

[0041] K (Key, key vector), dimension: , consistent with the dimension of Q, represents the "information tag carried by each position in the sequence" and is used to match Q and determine which position information is relevant to Q's "search intent." K at different times encodes labels such as "current pattern of wood drilling" and "torque fluctuation of metal drilling" for Q to filter.

[0042] V (Value) represents the actual content at each position in the sequence. When Q and K match and establish a correlation, V outputs the specific information of interest. For example, if Q matches K, which represents "metal material characteristics," V outputs specific data such as the current mutation amplitude and torque peak at that moment. The V dimension is aligned with Q and K (or adapted through linear transformation).

[0043] QKᵀ (query-key similarity calculation) calculates the "match score" for each position of Q and K through matrix multiplication. A higher value indicates a higher correlation between Q and K at that position. For example, if Q is "finding metal features" at a certain moment and the corresponding position K encodes "metal current pattern", the QKᵀ score at this position will be very high.

[0044] (scaling factor) is used to alleviate the problem of "softmax gradient disappearing due to too large a value" after QKᵀ calculation.

[0045] softmax(...) (normalization), The output "matching score" is converted into a probability distribution between 0 and 1 to ensure that the sum of the weights is 1, representing the "degree of attention paid to the information at each location."

[0046] softmax(...)V (weighted output), use the "attention weight" obtained by softmax to perform weighted summation on V (actual content), and output new features that integrate key information.

[0047] (3) Sequential pattern learning can be simplified into the formula: ;in, The feature sequence output by the feature extraction step is the local features initially extracted by 1D-CNN from the original data of the electric drill motor (current, torque, speed and other timing signals).

[0048] Transformer(...) (Transformer encoding process), based on the aforementioned encoder structure and self-attention mechanism The temporal relationship encoding model is constructed. Z is the temporal encoding vector; it is the global temporal feature vector output after Transformer encoding. The local features in the data are fused into a compact code that can represent the regularity of the entire time series data.

[0049] Specifically, as an example, the output of classification and state labels is based on a lightweight classifier, and the process can include: (1) Network structure: two fully connected layers: 64→32→C, where C=10 is the number of material types + state labels; Activation function: ReLU is used in the hidden layer and Softmax is used in the output layer.

[0050] (2) Classification formula: ; Among them, Z is the temporal coding vector obtained through the previous temporal pattern learning, is the weight matrix, which implements the preliminary weight adjustment and dimension conversion of the input feature Z to mine more discriminative relationships between features. Assuming that Z is 64-dimensional, Dimensions can be , realize the linear transformation from 64 dimensions to 32 dimensions; The bias term is used to adjust the result after linear transformation, improve the model fitting ability, and enable the model to learn more flexible feature representation; Activation functions, such as ReLU, introduce nonlinearity to the model, allowing it to learn complex feature relationships, avoiding simple linear combinations and improving its ability to fit complex operating scenarios and diverse material characteristics of electric drills.

[0051] The weight matrix can be , where C is the number of categories (for example, the material types include wood, metal, hollow brick, etc., and C is the number of these types). This achieves a linear transformation from 32 dimensions to C dimensions, further adjusting the feature weights to prepare for the final classification. This is the second-level bias term, which further adjusts the linear transformation results to help the model better fit the data and improve classification accuracy. Softmax converts the previous calculation results into a probability distribution, mapping the output values to the range 0-1, with the sum of all class probabilities equal to 1, indicating the likelihood that the current input belongs to each class.

[0052] y represents the probability distribution vector of each category output by Softmax, with a dimension of C (number of categories), and each element corresponds to the probability that the input belongs to a certain category.

[0053] It can be understood that the classifier can include two branches, namely the type classification branch and the state category branch. Through the fully connected layer mapping, the type branch and the state branch output the probability distribution of type and state respectively through softmax.

[0054] Taking type classification as an example, the softmax material type probability distribution can be calculated using the following formula: ;in, represents the conditional probability, which means the probability of belonging to the i-th class label (such as the i-th material type, state, etc.) under the condition of input data X (electric drill motor operation data, etc.); is the i-th category label, for example, the material type label can be "wood", "metal", "hollow brick", etc.; The input data is the collected motor operation data (time series data such as current, voltage, speed, torque, etc.).

[0055] The logits output of the i-th class label is the classification formula after After the operation, the original output value that has not been normalized by Softmax reflects the model's original score for the i-th category label. The larger the value, the more likely the model believes that the input belongs to this category.

[0056] C is the total number of categories, that is, the number of all possible label types. For example, if there are 5 material types, then C=5.

[0057] It is an exponential operation on the logits output of the i-th class label, The logits output of all category labels is summed after exponential operation, and the two are divided to get the probability of the i-th category label. This calculation can amplify The difference between them makes the probability distribution more focused.

[0058] In an optional embodiment, based on a preset association matrix of material type and material state, a probability distribution of material type and a probability distribution of material state, a joint probability distribution of material type and material state is calculated, and the material type and material state are determined based on the joint probability distribution.

[0059] Specifically, a type-state prior association rule matrix is established to form The dimensional association matrix R defines the compatibility of different material types and states. The type-state prior association rules are based on the empirical knowledge base constructed based on the physical properties of the material and historical data, which are used to correct the probability of independent predictions of the model to ensure that the joint prediction of the material type and material state conforms to the real logic. The association rules can be determined based on the physical properties of the material, historical data statistics or engineering practice experience. For example, wood has low hardness and is difficult to undergo hardening mutation. Metal has good ductility and is prone to "hardening mutation" due to friction and heating. In statistics of multiple drilling data, the probability of "metal + hardening mutation" is 35%, and "wood + hardening mutation" is only 2%. Due to the internal porous structure of hollow bricks, the probability of "void proximity" state is 40% higher than that of solid materials.

[0060] As an example, the material type : Wood, metal, hollow brick, ceramic, composite board, plastic; Material state Taking stability, hardening mutation, cavity proximity, and material mutation as examples, the association rule matrix example (partial) is shown in the following table:

[0061] The matrix elements It represents the prior probability correction coefficient of material type i being in state j, and the value range is [0,1.5]. It can be understood that the material type and material state can be set as needed, and this embodiment does not impose any restrictions.

[0062] As an example, for example, the independent prediction results are: type branch: P(metal) = 0.7, P(hollow brick) = 0.2; state branch: P(void proximity) = 0.6, P(hardening mutation) = 0.3. Combined with the above table, the joint probability correction based on the association rule: metal + void proximity: (Solid metal, low probability of voids), joint probability = .

[0063] Hollow brick + cavity approaching: (hollow brick inherent structure), joint probability = .

[0064] It can be seen that although the probability of "metal" is higher in the independent prediction, after the association rule is modified, the joint probability of "hollow brick + adjacent to the void" (0.18) is 4 times that of "metal + adjacent to the void" (0.042). Therefore, the final recognition result is that the type is identified as hollow brick and the status is adjacent to the void.

[0065] Step 103: Obtain the user's operational behavior data on the electric drill, wherein the operational behavior data includes the pressing duration, frequency, and rhythm changes of the electric drill trigger signal; based on the operational behavior data, obtain a second recognition result through a second deep learning model, and the second recognition result identifies the user's current usage intention, including tentative operation, continuous push operation, or impact operation.

[0066] Specifically, the user's operational behavior data on the electric drill is collected in real time through the built-in sensors and trigger switches of the electric drill. The trigger signal pressing duration represents the time interval from pressing to releasing the electric drill power switch (unit: ms). For example, the time difference between the on and off of the switch can be recorded by the timer module of the microcontroller, and the initial state of the motor at the pressing moment (such as the initial speed, the current torque setting value) can be synchronously marked; the trigger frequency and rhythm represent the number of triggers in a unit time (for example, 10s) (unit: times / 10s), and the distribution of time intervals between two adjacent triggers (for example, the shortest interval, the average interval). For example, the coefficient of variation (standard deviation / mean) of the trigger interval can be calculated by using sliding window statistics (window size 10s, step size 1s) to reflect the rhythm stability; the operation rhythm represents the combination mode of the trigger pressing duration and the speed control (for example, long press + high speed), and the correlation between the trigger frequency and torque regulation (for example, whether the torque increases synchronously when triggered at high frequency).

[0067] Exemplarily, the second deep learning model lightweight temporal neural network architecture includes three core modules: feature extraction layer: 1D-CNN feature extraction layer.

[0068] Used to capture local temporal features of operational behavior data (such as sudden changes in pressing time and short-term fluctuations in trigger frequency).

[0069] The structure is a 3-layer depthwise separable convolution (kernel size 5, 5, 3), with the number of channels increasing from 32 to 64 to 128, followed by BatchNorm and ReLU activation.

[0070] Time Series Modeling Layer: A bidirectional LSTM time series modeling layer is used to learn long-range dependencies in manipulation behaviors (e.g., the rhythmic pattern of consecutive presses and the changing trend of manipulation intensity). For example, this layer has a 128-dimensional hidden layer with bidirectional connections to capture both past and future time series information.

[0071] Classification layer: This layer uses an attention mechanism to focus on key operational features and outputs a probability distribution of intent. For example, the structure consists of an attention weight matrix, a fully connected layer, and a softmax layer, outputting the probabilities of three types of intent (exploratory, continuous, and aggressive).

[0072] As an example, the recognition process includes the following steps: 1. Data preprocessing: Timing standardization of the operation signal: The data of the trigger signal such as the pressing duration, frequency, and rhythm change are organized into a time series matrix. , where T=200 (2s window, sampling frequency 100Hz), D=3 (pressing time, triggering frequency, speed setting value).

[0073] Normalize the Z-Score of each dimension data, the formula is ,in, is the mean, is the standard deviation to ensure that the data distribution is stable.

[0074] 2. Local operation feature extraction: Through the feature extraction layer 1D-CNN layer, multi-scale features are captured through convolution kernels of different sizes: For example, the small kernel (3) identifies instantaneous changes in press time (e.g., press release within 100ms); the large kernel (5) captures short-term patterns in trigger frequency (e.g., high-frequency triggering of 2 times / second). For example, the local combination feature of "short press (<200ms) + high-frequency triggering" is extracted as a potential sign of tentative operation; the local feature of "long press (>500ms) + stable frequency" is extracted to indicate a continuous push operation.

[0075] 3. Long-Term Dependency Modeling: The bidirectional processing logic of the bidirectional LSTM layer is as follows: the forward LSTM learns the temporal sequence of actions from left to right (e.g., the process of first probing and then pushing); the backward LSTM captures the impact of future information on the present (e.g., the effect of a subsequent strong press on the current intention) from right to left. For example, for probing actions, the learning model learns the irregular repetitive pattern of "short press → release → short press"; for continuous pushing actions, the learning model learns the regular rhythm of "long press → stable interval → long press"; and for impact actions, the learning model learns the causal relationship of "long interval → single strong press → simultaneous increase in torque at high speed."

[0076] Update of hidden state: ,in is the hidden state at time t, For all previous operations, is the operation data at the current moment, and the two are fused to obtain the hidden state at moment t.

[0077] 4. Attention-based classification: Calculate the attention weight of each time step , allowing the model to focus on the key operation moments for intention judgment, which can be expressed by the formula: ;in, represents the attention weight at the t-th time step (0-1, the higher the weight, the more important), Indicates hidden state The relevance score to the query vector, measuring the value of judgments about intention); is the total number of time steps (e.g. 200 sampling points in a 2s window); exp is an exponential function that amplifies the difference in correlation scores (making important moments more prominent).

[0078] After the attention mechanism outputs the weights of each moment, a comprehensive feature vector is obtained through weighted summation, and then softmax separation is performed to compress the temporal features into intent probabilities. Specifically, the comprehensive feature vector can be calculated using the following formula: ; Among them, v represents the comprehensive feature vector, integrating the hidden states of all time steps , but the focus is weighted key moments through control, The hidden state at time t output by the bidirectional LSTM (including the operational features and historical dependencies at that moment).

[0079] The output of intent probability can be expressed as follows: .

[0080] W is the weight matrix of the fully connected layer (mapping the comprehensive feature v to the intent category dimension); b is the bias term (adjusting the classification boundary and improving model robustness). The softmax normalization function converts the output into a probability distribution between 0 and 1 (the sum of all intent probabilities is 1). The classification results may include tentative operations, continuous advancement operations, or impact operations. It is understood that the classification results can be trained and configured according to actual needs and are not limited in this embodiment.

[0081] Step 104: Based on the motor operation data, obtain a third recognition result through a third deep learning model, and the third recognition result includes identifying the drill bit status.

[0082] The third deep learning model focuses on identifying the health status of the drill bit. By analyzing the timing patterns of the motor operation data (current, torque, speed, etc.), it determines whether the drill bit is worn, blunted, broken, etc.

[0083] Specifically, the third deep learning model architecture can adopt the same architecture as the first deep learning model, such as a lightweight architecture of 1D-CNN+Transformer+attention classification: 1D-CNN layer: extract local abnormal features of current and torque (such as current spikes and torque mutations).

[0084] Transformer layer: Learns long-term temporal dependencies in motor data (e.g., the persistently high current pattern of a worn drill bit).

[0085] Attention classification layer: focuses on key failure moments (such as the torque drop point when the drill bit breaks) and outputs state probabilities.

[0086] The specific identification process will not be described in detail in this embodiment. It is understood that the third deep learning model can also adopt other lightweight models. The drill bit status can include normal, blunt, broken, stuck, etc., which can be set according to actual conditions and are not limited in this embodiment.

[0087] Step 105: Based on the first recognition result, the second recognition result, and the third recognition result, a control strategy for dynamically adjusting the torque is obtained; based on the control strategy, the output torque of the electric drill is dynamically adjusted.

[0088] By integrating three core pieces of information: material status (first recognition result), user intent (second recognition result), and drill bit status (third recognition result), a multi-dimensional decision-making model is constructed to generate a torque adjustment strategy that adapts to the current working conditions.

[0089] Specifically, the three types of recognition results are converted into standardized decision features to form a multi-dimensional input vector F. For example, the material state feature: for example, "metal + hardening mutation" is encoded as [1, 0, 0, 1] (type: metal = 1, state: hardening mutation = 1); the user intention feature: for example, "impact operation" is encoded as [0, 0, 1] (probe = 0, continuous = 0, impact = 1); the drill bit state feature: for example, "passivation" is encoded as [0, 1, 0, 0] (normal = 0, passivation = 1, fracture = 0, stuck drill = 0).

[0090] In an optional embodiment, a basic control strategy is determined based on a multi-dimensional input vector by querying a preset expert rule base. The expert rule base constructs a decision space based on three dimensions: material state (M), user intent (I), and drill bit state (B). For example, the basic control strategy of the expert rule base is shown in the following table:

[0091] In an optional embodiment, the first recognition result, the second recognition result and the third recognition result are encoded into a fused feature vector, and the fused feature vector is input into a pre-trained gradient boosting tree model to obtain a torque correction factor. The basic control strategy is dynamically weighted based on the torque correction factor, and the correction amount is accumulated over time by limiting the attenuation function to obtain the final torque adjustment strategy.

[0092] Specifically, dynamic weight correction uses the gradient boosting tree (GBT) to learn historical working condition data, achieving real-time fine-tuning of the basic strategy of the expert rule base, making torque adjustment more in line with the dynamic changes of actual operations.

[0093] The core of dynamic weight correction is to use historical experience to optimize current decisions. The process is as follows: 1. Data collection: record the full amount of data of electric drill operations (material status, user intention, drill bit status, torque strategy, motor feedback), and build a historical working condition database.

[0094] 2. Feature Engineering: Encode the three types of recognition results (material, intent, and drill bit) into a fused feature vector F, which includes: material mutation rate (for example, 2.5A / ms), user intent confidence (for example, 0.95 for impact operation), and drill bit blunting confidence (for example, 0.85).

[0095] 3. GBT Model Training: Using the matching degree between the torque strategy and actual motor feedback as the optimization objective, the GBT model is trained to learn the mapping relationship between F and the optimal torque correction. Specifically, during training, the model uses "optimal torque correction" as the target label and obtains the mapping relationship between feature F and this optimal torque correction (i.e., the function "feature F → optimal torque correction").

[0096] 4. Real-time correction: Input the current working condition F into the trained GBT model, and based on the mapping relationship between the characteristic F obtained during training and the optimal torque correction amount, output the correction factor based on the current working condition F , adjust the torque parameters of the basic strategy. Specifically, the correction factor Covering three types of dynamic features, acting on the core parameters of the torque strategy (initial torque, slope, delay), for example, (1) Material mutation intensity: Modify the trigger condition: material mutation rate > 2A / ms (for example, hard points suddenly appear in the metal layer).

[0097] Correction logic: when mutation rate > 2, increase torque , no correction is made in other cases.

[0098] It means that when the sudden change intensity is large, additional torque is required to break through the hard points to avoid the drill bit getting stuck.

[0099] (2) User intention strength: Modify the trigger condition: the confidence level of the impact operation is > 0.9 (the user explicitly performs a strong impact).

[0100] The correction logic is: when the confidence level is greater than 0.9, the delay is shortened by 50ms; otherwise, no correction is performed.

[0101] This means that when the user's intention is strong, the torque surge delay is shortened (for example, from 100ms to 50ms) to meet the "quick impact" requirement.

[0102] (3) Drill bit health correction: Trigger condition: Drill bit blunting confidence > 0.8 (drill bit cutting ability decreases).

[0103] Correction logic: When confidence > 0.8, speed up the torque increase slope , no correction is made in other cases.

[0104] This means that when the drill becomes blunt, the torque increase rate is accelerated (for example, from 0.5 to 0.7 N·m / s) to compensate for the loss in cutting force.

[0105] Optionally, a correction factor Integration method with basic strategy parameters: ;in, is the torque value actually executed by the electric drill, is the basic torque strategy of the aforementioned expert library, is a time decay function to avoid unlimited accumulation of corrections over time. For example,

[0106] Optionally, the dynamic adjustment of the output torque of the electric drill based on the control strategy includes analog output and dynamic response, wherein the analog output includes controlling the motor drive circuit through a PWM signal to adjust the torque output, and the dynamic response includes the control strategy taking effect delay time being less than a preset threshold.

[0107] Torque output is adjusted based on a control strategy (basic or modified). This includes analog output, which controls the motor drive circuit via PWM signals to adjust torque output (for example, 0-10V corresponds to 0-20N·m). Dynamic response, with a strategy activation delay of less than 50ms, meets the real-time demands of electric drilling operations (for example, impact operations require a sudden torque increase within 100ms).

[0108] like Figure 3 As shown, an embodiment of the present invention provides an electric drill torque dynamic adjustment device 300, comprising: a first acquisition module 301, configured to acquire motor operation data of the electric drill during operation, including at least time series data of current, voltage, control signal, speed, and motor load response; a first identification module 302, configured to obtain a first recognition result using a first deep learning model based on the motor operation data, wherein the first recognition result includes identifying the type and status of the drilled material; a second identification module 303, configured to obtain user operation behavior data on the electric drill, wherein the operation behavior data includes at least the pressing duration, frequency, and rhythm change of the electric drill trigger signal; and based on the operation behavior data, obtain a second recognition result using a second deep learning model, wherein the second recognition result identifies the user's current usage intention, including a tentative operation, a continuous push operation, or an impact operation; a third identification module 304, configured to obtain a third recognition result using a third deep learning model based on the motor operation data, wherein the third recognition result includes identifying the drill bit status; and an adjustment module 305, configured to obtain a control strategy for dynamically adjusting the torque based on the first, second, and third recognition results, and dynamically adjust the output torque of the electric drill based on the control strategy.

[0109] like Figure 4 As shown, an embodiment of the present invention provides an electronic device, including a processor and a memory, wherein the memory stores program instructions, and the processor runs the program instructions to implement the electric drill torque dynamic adjustment method as described in any one of the above items.

[0110] Another embodiment of the present application further provides a computer-readable storage medium having computer program instructions stored thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of the present application.

[0111] In particular, the methods and / or embodiments of the present application can be implemented as computer software programs. For example, the embodiments disclosed in the present application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for executing the method shown in the flowchart. When the computer program is executed by a processor, the above-mentioned functions defined in the method of the present application are performed.

[0112] It should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each implementation method can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0113] The series of detailed descriptions listed above are only specific descriptions of feasible implementation methods of the present invention. They are not intended to limit the scope of protection of the present invention. Any equivalent implementation methods or changes that do not deviate from the technical spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for dynamically adjusting torque of an electric drill, characterized in that: The following steps are involved: Obtaining motor operation data of the electric drill during operation, including at least time series data of current, voltage, control signal, speed, and motor load response; Based on the motor operation data, a first recognition result is obtained using a first deep learning model, wherein the first recognition result includes identifying the type and status of the drilled material; user operation behavior data on the electric drill is obtained, wherein the operation behavior data at least includes the pressing duration, frequency, and rhythm changes of the electric drill trigger signal; based on the operation behavior data, a second recognition result is obtained using a second deep learning model, wherein the second recognition result identifies the user's current usage intention, including a tentative operation, a continuous push operation, or an impact operation; Based on the motor operation data, a third recognition result is obtained through a third deep learning model, and the third recognition result includes identifying the drill bit status; based on the first recognition result, the second recognition result and the third recognition result, a control strategy for dynamically adjusting the torque is obtained, and based on the control strategy, the output torque of the electric drill is dynamically adjusted.

2. The method for dynamic torque adjustment of an electric drill according to claim 1, characterized in that: include: The first recognition result is obtained through the first deep learning model, including calculating the joint probability distribution of the material type and the material state based on a preset association matrix of the material type and the material state, the probability distribution of the material type, and the probability distribution of the material state, and determining the material type and the material state based on the joint probability distribution.

3. The method for dynamic torque adjustment of an electric drill according to claim 1, characterized in that: include: The obtaining of the control strategy for dynamically adjusting the torque based on the first recognition result, the second recognition result and the third recognition result includes querying a preset expert rule library based on the first recognition result, the second recognition result and the third recognition result to obtain a basic control strategy.

4. The method for dynamically adjusting the torque of an electric drill according to claim 3, characterized in that: include: The first recognition result, the second recognition result, and the third recognition result are encoded into a fused feature vector, which is input into a pre-trained gradient boosting tree model to obtain a torque correction factor. The basic control strategy is dynamically weighted based on the torque correction factor, and the accumulation of the correction amount is limited by a time decay function to obtain the final torque regulation control strategy.

5. The method for dynamic torque adjustment of an electric drill according to claim 4, characterized in that: include: The first deep learning model adopts a lightweight time series deep learning architecture, which is optimized for the time series characteristics of electric drill motor data, and includes a first feature extraction layer, a first time series modeling layer, and a first classification prediction layer; wherein the first feature extraction layer is a 1D-CNN for capturing the local features of the motor operation data; the first time series modeling layer is a Transformer encoder or an LSTM unit for learning long-distance dependencies and outputting a global time series feature vector; the first classification prediction layer includes a fully connected layer and a classification function for outputting the first recognition result.

6. The method for dynamically adjusting the torque of an electric drill according to claim 1, characterized in that: include: The dynamic adjustment of the output torque of the electric drill based on the control strategy includes analog output and dynamic response, wherein the analog output includes controlling the motor drive circuit through a PWM signal to adjust the torque output, and the dynamic response includes the control strategy taking effect delay time being less than a preset threshold.

7. A dynamic torque adjustment device for an electric drill, characterized in that: include: The first acquisition module is used to obtain the motor operation data of the electric drill during operation, including at least the time series data of the current, voltage, control signal, speed and motor load response; a first recognition module, configured to obtain a first recognition result by using a first deep learning model based on the motor operation data, wherein the first recognition result includes identifying the type and status of the drilled material; A second recognition module is configured to obtain user operational behavior data on the electric drill, the operational behavior data including at least the duration, frequency, and rhythm changes of the press of the electric drill trigger signal, and obtain a second recognition result based on the operational behavior data using a second deep learning model. The second recognition result identifies the user's current usage intention, including a tentative operation, a continuous push operation, or an impact operation; a third recognition module, configured to obtain a third recognition result based on the motor operation data by using a third deep learning model, wherein the third recognition result includes identifying a drill bit status; The adjustment module is used to obtain a control strategy for dynamically adjusting the torque based on the first recognition result, the second recognition result and the third recognition result, and dynamically adjust the output torque of the electric drill based on the control strategy.

8. An electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, characterized in that the instructions are executed by the at least one processor so that the at least one processor can execute the method according to any one of claims 1 to 6.

9. A computer-readable medium having computer program instructions stored thereon, characterized in that: The computer program instructions can be executed by a processor to implement the method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images