Intelligent manufacturing abnormal state recognition method and system based on artificial intelligence

Abstract

The application provides an intelligent manufacturing abnormal state recognition method and system based on artificial intelligence, which comprises the following steps: acquiring a physical quantity monitoring signal; constructing an abnormal feature extraction network embedded with a physical mechanism; physically initializing the convolution kernel weight of the first convolution layer by using an impact response function corresponding to the dynamic behavior of equipment failure, to generate a physically guided fault sensitive feature map; determining a non-uniform downsampling rate according to a fault feature frequency ratio, performing frequency adaptive compression on the feature map, and generating a fault feature representation with reduced dimensions; a classification output layer performs classification mapping on the compressed features, and outputs the confidence scores of various states; during training, a loss function containing a physical consistency regularization term is used to constrain the convolution kernel spectrum energy to be concentrated at the fault feature frequency and its multiple frequencies; and during inference, real-time signals are input into the trained network to obtain an abnormal state recognition result. The method can effectively improve the accuracy and physical interpretability of abnormal state recognition under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing, and in particular to an artificial intelligence-based method and system for identifying abnormal states in intelligent manufacturing. Background Technology

[0002] In intelligent manufacturing anomaly identification, existing technologies typically employ data-driven deep neural networks to directly process physical quantity monitoring signals. Models such as convolutional neural networks automatically extract features from these signals and output classification results. The network's convolutional kernel weights are initialized randomly and trained end-to-end using a large number of labeled samples. This random initialization method results in a lack of awareness of the physical laws governing equipment failures during the initial training phase. When network training relies entirely on data feedback, and failure samples are scarce or imbalanced, the feature extraction direction learned by the network tends to deviate from the essential characteristics of the failure, tending to fit noise or non-fault-related patterns in the training data. This leads to insufficient stability and consistency of anomaly identification results under complex industrial conditions. Summary of the Invention

[0003] This invention provides a method and system for identifying abnormal states in intelligent manufacturing based on artificial intelligence.

[0004] In a first aspect, embodiments of the present invention provide an intelligent manufacturing anomaly state identification method based on artificial intelligence, including: Acquire physical quantity monitoring signals collected by industrial equipment during operation; An anomaly feature extraction network embedded with physical mechanisms is obtained. The anomaly feature extraction network embedded with physical mechanisms includes a first convolutional layer, a physical pooling layer, and a classification output layer connected in sequence. The first convolutional layer is physically initialized. The initial values ​​of the convolutional kernel weights of the first convolutional layer are assigned using the impact response function corresponding to the dynamic behavior of equipment faults. This makes the forward propagation process of the first convolutional layer equivalent to performing a fault feature transformation under physical constraints on the physical quantity monitoring signal, generating a physically guided fault-sensitive feature map. The physically guided fault-sensitive feature map is input into the physical pooling layer. The physical pooling layer determines the non-uniform downsampling rate according to the preset fault feature frequency ratio, and uses the non-uniform downsampling rate to perform frequency-adaptive feature compression processing on the physically guided fault-sensitive feature map to generate a fault feature representation with reduced dimensionality. The reduced-dimensional fault feature representation is input into the classification output layer, which performs classification mapping on the reduced-dimensional fault feature representation to obtain the confidence score of whether the industrial equipment belongs to the normal state or various abnormal states. The anomaly feature extraction network embedded with physical mechanisms is trained using a loss function that includes a physical consistency regularization term. The physical consistency regularization term is used to constrain the spectral energy distribution of the convolution kernel of the first convolutional layer to be concentrated at the preset fault feature frequency and its harmonics in the frequency domain. During the inference phase, the real-time collected physical quantity monitoring signal is input into the trained anomaly feature extraction network embedded with physical mechanisms, and then processed sequentially through the first convolutional layer, the physical pooling layer, and the classification output layer to obtain the anomaly state recognition result.

[0005] Secondly, embodiments of the present invention provide a computer system, including: a memory for storing computer-executable instructions or computer programs; and a processor for executing the computer-executable instructions or computer programs stored in the memory to implement the above-mentioned artificial intelligence-based intelligent manufacturing abnormal state identification method.

[0006] The embodiments of the present invention have the following beneficial effects: This invention embeds the impact response function corresponding to the dynamic behavior of industrial equipment failure directly into the first convolutional layer of an anomaly feature extraction network, assigning weights to the convolutional kernels through physical initialization. This enables the convolutional operation itself to possess the ability to match fault features under physical constraints, allowing the network to extract fault-sensitive feature maps with clear physical meaning from physical quantity monitoring signals without relying on a large number of fault samples. Subsequently, a physical pooling layer performs non-uniform frequency adaptive compression based on the fault feature frequency ratio, ensuring that the feature structure related to the fault impact is finely preserved, while features in stable regions unrelated to the fault are significantly compressed. This reduces feature dimensionality while suppressing background interference from confusing subsequent classifications. A physical consistency regularization term introduced during training continuously constrains the spectral energy of the convolutional kernels to concentrate at the fault feature frequencies and their harmonics, ensuring that the network learning process always follows the physical prior of the equipment fault rather than merely fitting the data distribution. This alleviates the overfitting tendency common in data-driven methods and endows the network parameters with interpretable physical meaning. During the inference phase, the real-time monitoring signal flows sequentially through the physically guided convolutional layer, frequency-adaptive pooling layer, and classification output layer. The entire process does not require manual threshold setting or manual feature construction, and can output the confidence score of the device belonging to each state category. It can still maintain stable anomaly recognition capability under conditions of few samples. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the architecture of an application scenario provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the computer system provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the intelligent manufacturing anomaly state identification method based on artificial intelligence provided in this embodiment of the invention.

[0008] Figure 4 This is a flowchart illustrating the intelligent manufacturing anomaly state identification method based on artificial intelligence provided in this embodiment of the invention. Detailed Implementation

[0009] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0010] See Figure 1 This is an architectural diagram of an application scenario provided in an embodiment of the present invention. Industrial equipment 400 is connected to computer system 200 through network 300. Network 300 can be a wide area network or a local area network, or a combination of the two, and data transmission is achieved using wireless or wired links.

[0011] In some embodiments, the computer system 200 may be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Industrial equipment 400 may be, for example, an induction motor driving a compressor, an asynchronous motor for a pump, a wind turbine main bearing housing, a belt conveyor drive roller bearing, a rolling mill gearbox, and a turbocharger rotor system.

[0012] The computer system implementing the AI-based intelligent manufacturing anomaly identification method provided in this embodiment of the invention will be described next. See [link to documentation]. Figure 2 , Figure 2 This is a schematic diagram of the structure of the computer system provided in an embodiment of the present invention. Figure 2 The computer system shown includes at least one processor 210, memory 250, at least one network interface 220, and an external interface 230. The various components in the computer system 200 are coupled together via a bus system 240. It is understood that the bus system 240 is used to implement communication between these components. In addition to a data bus, the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 2 The general labeled all buses as Bus System 240.

[0013] Processor 210 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0014] External interface 230 may include, for example, one or more speakers and / or one or more visual displays. External interface 230 may also include one or more input devices 232, such as a keyboard, mouse, microphone, touch screen display, camera, etc.

[0015] The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 250 may optionally include one or more storage devices physically located away from the processor 210.

[0016] The memory 250 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 250 described in this embodiment is intended to include any suitable type of memory.

[0017] In some embodiments, memory 250 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0018] Operating system 251 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks; The network communication module 252 is used to reach industrial equipment via one or more (wired or wireless) network interfaces 220, exemplary network interfaces 220 including: Bluetooth, WiFi, and Universal Serial Bus (USB), etc. Presentation module 253 is configured to enable the display of information (e.g., external interface for operating peripheral devices and displaying content and information) via one or more output devices 231 (e.g., display screen, speaker, etc.) associated with external interface 230; The input processing module 254 is used to detect and translate one or more user inputs or interactions from one or more input devices 232.

[0019] Based on the above description of the application scenarios and computer systems provided in the embodiments of the present invention, the following describes the artificial intelligence-based intelligent manufacturing anomaly state identification method provided in the embodiments of the present invention. In actual implementation, the artificial intelligence-based intelligent manufacturing anomaly state identification method provided in the embodiments of the present invention can be implemented by a computer system. Combined with... Figure 3 and Figure 4 Referring to the present invention, the method for identifying abnormal states in intelligent manufacturing based on artificial intelligence will be described below, specifically including: Step S100: Acquire physical quantity monitoring signals collected by industrial equipment during operation.

[0020] Industrial equipment refers to a collection of mechanical devices that perform specific technological actions in a production and manufacturing environment. Its specific forms can include rotary power machinery, reciprocating motion machinery, fluid transport machinery, or thermodynamic conversion machinery. In intelligent manufacturing scenarios, possible examples include harmonic reducers inside the joints of industrial robots, electric spindle units in high-speed CNC machining centers, planetary gear speed increasers in wind turbine generators, and multi-stage centrifugal compressors in compressed air energy storage systems. Physical quantity monitoring signals are time-varying sequences of physical quantity values ​​measured from the surface or interior of industrial equipment through sensing and conversion devices. These signals reflect the mechanical response state generated by the interaction between various components during equipment operation. The sensing sources for physical quantity monitoring signals can include vibration acceleration sensors, dynamic displacement sensors, acoustic emission sensors, dynamic torque sensors, and temperature field sensors. These sensors continuously acquire signals at a fixed sampling frequency, converting the amplitude changes of physical quantities into a discrete numerical sequence arranged in time order.

[0021] For example, in the condition monitoring scenario of an industrial robot wrist joint reducer, a piezoelectric vibration acceleration sensor mounted on the reducer's flex wheel housing collects the vibration acceleration signal transmitted to the housing surface when the flex wheel meshes with the rigid wheel inside the reducer at a fixed sampling frequency. This vibration acceleration signal constitutes a set of physical quantity monitoring signals. As another example, in the condition monitoring scenario of a high-speed CNC machining center's electric spindle, an eddy current displacement sensor mounted on the front and rear bearing seats of the spindle collects the spindle centerline trajectory offset signal caused by changes in bearing clearance during spindle rotation at a fixed sampling frequency. This spindle centerline trajectory offset signal also constitutes a physical quantity monitoring signal. During normal equipment operation, the physical quantity monitoring signal manifests as a stable and continuous background vibration component, with its amplitude statistical distribution within a relatively concentrated range. However, when abnormalities such as localized damage, assembly misalignment, or lubrication deterioration occur inside the equipment, transient impact response components will intermittently appear in the signal. The waveform shape, repetition frequency, and attenuation characteristics of these impact response components have a definite dynamic mapping relationship with the equipment's fault type, fault location, and operating conditions.

[0022] Step S200: Obtain the anomaly feature extraction network embedded with physical mechanism. The anomaly feature extraction network embedded with physical mechanism includes a first convolutional layer, a physical pooling layer and a classification output layer connected in sequence. Perform physical initialization processing on the first convolutional layer. Use the impact response function corresponding to the dynamic behavior of equipment faults to assign initial values ​​to the convolutional kernel weights of the first convolutional layer, so that the forward propagation process of the first convolutional layer is equivalent to performing fault feature transformation under physical constraints on the physical quantity monitoring signal, and generating a physically guided fault sensitive feature map.

[0023] The anomaly feature extraction network embedded with physical mechanisms is an artificial neural network model that deeply integrates prior knowledge of industrial equipment fault dynamics into its internal structure and operational logic. This network comprises three sequentially connected hierarchical modules: a first convolutional layer, a physical pooling layer, and a classification output layer. The output of the previous layer serves as the input to the next layer. The first convolutional layer is the initial processing layer of the network. Its structure includes a set of convolutional kernels, each of which is a one-dimensional weight vector. The length of the weight vector is equal to the preset kernel length, and the number of kernels is equal to the preset channel depth. During operation, the first convolutional layer performs a sliding window cross-correlation operation on the input physical quantity monitoring signal. At each sliding position, it calculates the dot product between the kernel weights and the local signal segment, generating a time-varying sequence of response values. The physical pooling layer, located between the first convolutional layer and the classification output layer, is a non-uniform downsampling module. This module receives the physically guided fault-sensitive feature map output from the first convolutional layer and performs downsampling operations with different compression ratios on different time segments of the feature map based on the physical frequency information of repeated fault impacts, outputting a feature representation with reduced time dimension. The classification output layer, following the physical pooling layer, consists of a fully connected mapping layer and a flexible maximum transfer function output layer. The fully connected mapping layer flattens the feature representation output from the physical pooling layer and maps it to the state class space. The flexible maximum transfer function output layer transforms the mapping result into a normalized confidence probability distribution. Physical initialization is an operation that assigns initial parameters to the network using analytical knowledge of the device fault dynamics. This operation is performed before the start of network training. Its purpose is to ensure that the convolutional kernel weights of the first convolutional layer naturally possess morphological characteristics matching the fault impact response waveform in the initial state, thereby obtaining sensitivity to fault impact signals at the beginning of training, accelerating the training convergence process, and improving the physical interpretability of the network. The impact response function, corresponding to the dynamic behavior of equipment failure, is a mathematical function describing the transient oscillatory response of industrial equipment's mechanical structure when subjected to a failure impact excitation, which gradually decays under structural damping. The specific form of this function is determined by the equipment's natural frequency of the impact response and the damping decay trend. In the time domain, it presents as a waveform that rapidly rises to a peak value from the initial impact moment and then decays exponentially with an envelope. The physically guided fault-sensitive feature map is the output obtained by performing a convolution operation on the input physical quantity monitoring signal after the first convolutional layer completes the physical initialization process. This feature map is a two-dimensional data structure. Its first dimension is the time direction, with a length equal to the output length after the convolution operation of the input signal length. The second dimension is the channel direction, with a depth equal to the number of convolution kernels in the first convolutional layer. Each channel in the feature map independently reflects the matching response intensity of the impact response template carried by the convolution kernel of that channel.

[0024] In one implementation, step S200 may specifically include the following steps S210 to S260: Step S210: Analyze the inherent structural indication information of the industrial equipment, and separate the operating frequency indication information and geometric dimension indication information that affect the dynamic response mode of the fault impact. The operating frequency indication information represents the periodicity of the rotation of the main shaft of the equipment, and the geometric dimension indication information represents the diameter of the rolling element distribution circle, the individual diameter of the rolling element, the total number of rolling elements, and the contact tilt angle.

[0025] The inherent structural indication information of industrial equipment is a collection of technical data related to the equipment's design parameters and assembly specifications. Its sources can include nameplate parameters on the equipment body, dimension markings in mechanical assembly drawings, numerical fields in parts specification tables, or performance parameter sections in the technical manuals provided by the equipment manufacturer. Operating frequency indication information describes the periodic speed of the equipment's spindle rotation. Its physical meaning is the number of complete rotations the spindle completes per unit time. This information influences the periodic arrangement of the impact response waveform by determining the macroscopic time interval of the failure impact. Geometric dimension indication information is a collection of parameters describing the spatial distribution and geometry of the rolling elements within the equipment. The diameter of the rolling element distribution circle refers to the diameter of the circumference of the circle containing the centers of all rolling elements; the individual diameter of a rolling element refers to the outer diameter of a single rolling element; the total number of rolling elements refers to the number of rolling elements installed inside the equipment; and the contact tilt angle refers to the angle by which the contact direction between the rolling element and the raceway deviates from the radial plane. In this embodiment, the model identification string of the industrial equipment is read from its inherent structural indication information. Using this model identification string as the search key, an exact match query is performed in a pre-built equipment parameter database. This database is a relational data table, and its fields include model identification, rated speed, pitch circle diameter, rolling element diameter, number of rolling elements, and contact angle. Upon a match, the value of the rated speed field is extracted from the corresponding record as the operating frequency indication information; the value of the pitch circle diameter field is extracted as the diameter of the rolling element distribution circle; the value of the rolling element diameter field is extracted as the individual diameter of the rolling element; the integer value of the number of rolling elements field is extracted as the total number of rolling elements; and the value of the contact angle field is extracted as the contact tilt angle.

[0026] Step S220: Using the operating frequency indication information and geometric dimension indication information, drive the fault impact response generation process to generate a damped oscillation waveform sequence simulating a single fault impact. The damped oscillation waveform sequence is composed of the rapid rising edge of the impact oscillation initiation stage and the exponential envelope falling edge of the damping decay stage.

[0027] In one implementation, step S220 may specifically include the following steps S221 to S226: Step S221: Generate repetition frequency indication information of the occurrence of fault impact events from the operating frequency indication information and geometric dimension indication information, and use the repetition frequency indication information as the fault characteristic frequency, which represents the number of times the fault impact occurs per unit time; obtain the impact response natural frequency indication information of the industrial equipment, which reflects the resonance frequency of the equipment structure under fault impact excitation.

[0028] In one implementation, step S221 may specifically include the following steps S2211 to S2216: Step S2211: Separate the rolling element pitch circle diameter indication information, rolling element diameter indication information, rolling element quantity indication information, and contact angle indication information from the geometric dimension indication information. The rolling element pitch circle diameter indication information is the diameter of the center distribution circle of the rolling element, the rolling element diameter indication information is the diameter of a single rolling element, and the contact angle indication information is the angle between the contact direction of the rolling element and the raceway and the radial plane.

[0029] In this embodiment, the geometric dimension indication information obtained in step S210 is further refined and extracted into various sub-parameters: the value of the pitch circle diameter field is read as the rolling element pitch circle diameter indication information, which represents the diameter of the circumference where the centers of all rolling elements are located; the value of the rolling element diameter field is read as the rolling element diameter indication information, which represents the outer diameter of a single rolling element; the integer value of the rolling element number field is read as the rolling element number indication information, which represents the total number of rolling elements installed inside the bearing; and the value of the contact angle field is read as the contact angle indication information, which represents the angle between the rolling element and the inner or outer ring raceway contact line deviating from the radial plane. These four sub-parameters are the basic geometric input quantities for calculating the inner ring impact characteristic frequency, the outer ring impact characteristic frequency, and the rolling element impact characteristic frequency.

[0030] Step S2212: Based on the rolling element pitch circle diameter indication information, rolling element diameter indication information, and contact angle indication information, and combined with the operating frequency indication information, generate the inner ring impact characteristic frequency of the rolling element passing through the inner ring fault point, the outer ring impact characteristic frequency of the rolling element passing through the outer ring fault point, and the rolling element impact characteristic frequency caused by the rolling element's own defects.

[0031] To generate the inner ring impact characteristic frequency, first calculate the ratio of the difference between the rolling element pitch circle diameter and the rolling element diameter multiplied by the cosine of the contact angle, to the rolling element pitch circle diameter. Add 1 to this ratio, multiply by the operating frequency indication, then multiply by the rolling element quantity indication and divide by 2. The resulting value is the inner ring impact characteristic frequency, in Hertz. The outer ring impact characteristic frequency can be generated by first calculating the ratio of the sum of the rolling element pitch circle diameter and the rolling element diameter multiplied by the cosine of the contact angle, to the rolling element pitch circle diameter. Subtract 1 from this ratio, take the absolute value, multiply by the operating frequency indication, then multiply by the rolling element quantity indication and divide by 2. The resulting value is the outer ring impact characteristic frequency, also in Hertz. The rolling element impact characteristic frequency is generated by first calculating the ratio of the rolling element pitch circle diameter to the rolling element diameter, then multiplying this ratio by the operating frequency indication to obtain an intermediate product. Then, calculate the square of the rolling element diameter, divide it by the square of the rolling element pitch circle diameter, and multiply by the square of the contact angle cosine to obtain a proportional value. Subtract this proportional value from the numerical value to obtain a correction factor. Finally, multiply the intermediate product result by this correction factor to obtain the rolling element impact characteristic frequency, in Hertz.

[0032] Step S2213: The inner ring impact characteristic frequency, outer ring impact characteristic frequency and rolling element impact characteristic frequency are aggregated into a fault impact event frequency set, which includes the repetition frequency corresponding to at least one fault type.

[0033] The set of failure impact event frequencies is a data structure containing three values. The first element is the inner ring impact characteristic frequency generated in step S2212, the second element is the outer ring impact characteristic frequency generated in step S2212, and the third element is the rolling element impact characteristic frequency generated in step S2212. This set covers the impact repetition frequency of each of the three typical failure locations of rolling bearings.

[0034] Step S2214: Obtain the fault type indication information recorded in the historical fault records of industrial equipment, and select a specific fault impact frequency that is consistent with the fault type indication information from the fault impact event frequency set as the preferred repetition frequency.

[0035] Historical fault records are collections of fault event documents accumulated during the past operation and maintenance of industrial equipment. Each record contains at least the fault occurrence time, fault location type, and fault description fields. The fault type indication information is a data identifier extracted from historical fault records that indicates the type of fault location that has historically occurred on the equipment. For example, the value of this identifier can be an inner race fault type, an outer race fault type, or a rolling element fault type. In this embodiment, the system reads the most recent fault record from the historical fault records and extracts the fault type indication information from its fault location type field. If the indication information indicates an inner race fault type, the inner race impact characteristic frequency is selected as the preferred repetition frequency from the fault impact event frequency set; if the indication information indicates an outer race fault type, the outer race impact characteristic frequency is selected; if the indication information indicates a rolling element fault type, the rolling element impact characteristic frequency is selected. This historical record-based optimization strategy ensures that the repetition period of the generated impact response template sequence more closely matches the actual fault tendency of the equipment.

[0036] Step S2215: When historical fault records are missing, extract each candidate impact frequency from the fault impact event frequency set in sequence, drive subsequent steps to generate their respective fault sensitive feature maps, and obtain the classification mapping response intensity of the classification output layer to each fault sensitive feature map. Determine the retention order of candidate impact frequencies based on the response intensity.

[0037] When there are no historical fault records available for reference, it is impossible to directly determine which fault characteristic frequency in the fault impact event frequency set should be prioritized to drive physical initialization. In this case, a trial-and-error evaluation strategy is adopted: for the three candidate impact frequencies in the fault impact event frequency set—inner race impact characteristic frequency, outer race impact characteristic frequency, and rolling element impact characteristic frequency—the complete physical initialization process of subsequent steps S222 to S260 is executed independently and sequentially, generating three different sets of first convolutional layer convolutional kernel weights. Each set of weights is assigned a value based on the impact response template sequence generated by the corresponding candidate impact frequency. For the three sets of convolutional kernel weight versions embedded with the physical mechanism, the abnormal feature extraction network performs forward propagation calculations using a labeled small-scale validation signal set, obtaining the confidence score for each category of each validation sample at the output of the classification output layer. The average confidence score of all real fault samples in the validation signal set for their actual fault categories is taken as the classification mapping response intensity of the corresponding version of the convolutional kernel. Compare the corresponding classification mapping response intensity values ​​of the three candidate impact frequency versions, and sort the three candidate impact frequencies in descending order of response intensity. The candidate impact frequency with the highest response intensity is ranked first. This frequency is most likely to stimulate an effective response to the real fault signal during physical initialization and best matches the current actual operating state of the equipment.

[0038] Step S2216: Output the final retained fault impact frequency as the repetition frequency information of the fault impact event. The repetition frequency information is used to determine the splicing period of the subsequent impact response template sequence.

[0039] When historical fault records exist, the final retained fault impact frequency is the specific fault impact frequency selected from the fault impact event frequency set in step S2214 that corresponds to the fault type pointing information. When historical fault records are missing, the final retained fault impact frequency is the candidate impact frequency with the highest classification mapping response intensity determined through tentative evaluation in step S2215. This final retained fault impact frequency is output in numerical form and passed to step S230 to determine the fixed time interval, i.e., the splicing period, between each impact event in the impact response template sequence.

[0040] Step S222: Using the natural frequency of the impact response as the oscillation frequency, generate a sinusoidal oscillation fundamental wave with the same periodicity as the natural frequency of the impact response. The sinusoidal oscillation fundamental wave is a continuous periodic waveform with stable amplitude, and its period duration is the reciprocal of the natural frequency of the impact response.

[0041] The fundamental sinusoidal oscillation is a discrete-time sequence of a standard sinusoidal waveform whose oscillation frequency equals the natural frequency of the impact response and whose amplitude does not decay over time. In this embodiment, the fundamental sinusoidal oscillation is generated as follows: using the natural frequency of the impact response as a parameter, the angular frequency of the sinusoidal waveform is determined to be twice the natural frequency of the impact response multiplied by the natural frequency of the impact response. The sampling time interval is determined according to the sampling frequency of the physical quantity monitoring signal, and the sampling time interval is the reciprocal of the sampling frequency. Starting from time zero on the time axis, a series of discrete time points are generated with the sampling time interval as the step size. For each discrete time point, the corresponding sine function value is calculated. The phase parameter of the sine function is initialized to zero, and the function amplitude parameter is set to unit amplitude. The sine function values ​​corresponding to all discrete time points are arranged in chronological order to form the fundamental sinusoidal oscillation sequence. The number of sampling points contained in each oscillation period of this sequence is equal to the quotient of the sampling frequency divided by the natural frequency of the impact response, and the duration of the oscillation period is the reciprocal of the natural frequency of the impact response.

[0042] Step S223: Obtain damping attenuation trend indication information reflecting the damping characteristics of the industrial equipment structure. The damping attenuation trend indication information is determined by the material internal friction and the friction characteristics of the joint surface. Use the damping attenuation trend indication information to generate an exponential attenuation envelope. The exponential attenuation envelope shows a decreasing trend in the time domain, gradually approaching zero from the initial time to subsequent time.

[0043] Damping attenuation trend indication information is a physical parameter describing how quickly the vibration amplitude of a device's mechanical structure decays over time in a free oscillation state. Specifically, it can be characterized by the damping ratio or the attenuation time constant. The damping ratio is defined as the ratio of the actual damping coefficient of the structure to the critical damping coefficient. The attenuation time constant is defined as the time required for the vibration amplitude to decay from its initial value to one-tenth of the base of the natural logarithm of the initial value. These two can be converted to each other using the natural frequency of the impact response. In this embodiment, the exponential attenuation envelope is generated as follows: using the base of the natural logarithm as the base, the negative value of the product of the damping ratio corresponding to the damping attenuation trend indication information and the natural frequency of the impact response as the exponential coefficient, and using the sampling time sequence on the time axis as the exponential variable, the corresponding power of the base of the natural logarithm is calculated point by point to generate a sequence that starts from a unit amplitude at the initial moment, monotonically decreases with time, and gradually approaches 0. This sequence is the exponential attenuation envelope. The sampling time interval used for the exponential attenuation envelope is consistent with the sampling time interval used in step S222 when generating the sinusoidal oscillation fundamental wave, to ensure that the sampling points of both are aligned point by point on the time axis.

[0044] Step S224: Modulate the instantaneous amplitude of the sinusoidal oscillation fundamental wave with the corresponding amplitude of the exponentially decaying envelope point by point, so that the oscillation amplitude of the sinusoidal oscillation fundamental wave changes synchronously according to the decay trend of the exponentially decaying envelope, thereby generating a decaying oscillation waveform of a single impact.

[0045] Point-by-point modulation is an operation that multiplies the values ​​at the same index position of two time series of the same length point by point. In this embodiment, point-by-point traversal is performed on the sinusoidal oscillation fundamental wave sequence and the exponentially decaying envelope sequence: for each sampling point index position, the instantaneous amplitude of the sinusoidal oscillation fundamental wave at that position is read, the envelope amplitude of the exponentially decaying envelope at that position is read, and the two are multiplied. The product is used as the amplitude of the decaying oscillation waveform at that sampling point index position. After traversing all sampling points, the resulting sequence is the decaying oscillation waveform. The internal oscillation frequency of this waveform always remains at the natural frequency of the impact response, but its peak envelope of oscillation is completely modulated by the exponentially decaying envelope, exhibiting a physical decay characteristic that gradually decays from the initial impact moment to near zero.

[0046] Step S225: Determine the position of the transient response peak at the start of the impact in the damped oscillation waveform, and take the transient response peak position as the starting point to intercept the pre-impact stable segment. The waveform amplitude in the pre-impact stable segment is stable relative to the amplitude at the transient response peak position.

[0047] The transient response peak position refers to the index position corresponding to the sampling point where the absolute value of the amplitude reaches its maximum in the decaying oscillation waveform sequence. In this embodiment, the operation method for determining the peak position is as follows: traverse each sampling point in the decaying oscillation waveform sequence, record its absolute value of amplitude, track the maximum absolute value and its corresponding sampling point index during the traversal process, and the index position obtained at the end of the traversal is the transient response peak position. The pre-impact stable segment is a waveform segment that simulates the stable operating background before the transient response peak position. In this embodiment, the pre-impact stable segment is generated by: tracing back a sampling point interval of a preset length from the transient response peak position. This preset length is determined according to the total leading stable segment duration required when designing the impact response template sequence. The amplitude of each sampling point in the traversal interval is set to zero, or set to a small amplitude random noise value centered at zero, so that the overall amplitude in this interval presents a low amplitude state that is significantly stable relative to the transient response peak position. This pre-impact stable segment simulates the signal pattern of the equipment in steady-state operation before the fault impact occurs.

[0048] Step S226: The pre-impact stable segment and the damped oscillation waveform are spliced ​​together end to end. The end of the pre-impact stable segment is connected to the beginning of the damped oscillation waveform to form a complete single fault impact damped oscillation waveform sequence that includes the stable state before impact and the damped oscillation state after impact.

[0049] End-to-end splicing is the operation of connecting two time series into a continuous time series in chronological order. In this embodiment, an empty output sequence is created. First, all sampling points of the pre-impact steady-state segment sequence are appended to the output sequence in their original order, and the index position of the last sampling point of the pre-impact steady-state segment is recorded. Then, starting from the first sampling point of the damped oscillation waveform sequence, all sampling points of that sequence are appended to the output sequence after the last sampling point in their original order. The output sequence obtained after splicing is the single-fault impact damped oscillation waveform sequence. The first part of this sequence represents the steady state before the impact, the middle part represents the rapid rising edge and transient response peak, and the last part represents exponentially damped oscillation. This complete waveform sequence fully reproduces the complete physical process on the time axis from the steady state before the impact to the instant of the impact and then to the gradual decay after the impact.

[0050] Step S230: Obtain the fault impact repetition frequency generated by the operating frequency indication information and the geometric dimension indication information, and cyclically copy and splice the damped oscillation waveform sequence according to the impact repetition period determined by the fault impact repetition frequency. Maintain an interval consistent with the impact repetition period between the end of the previous waveform sequence and the beginning of the next waveform sequence to form an impact response template sequence with continuous impact interval characteristics.

[0051] The fault impact repetition frequency is the final retained fault impact frequency output in step S2216. The reciprocal of this frequency is the impact repetition period, which represents the average time interval between two adjacent fault impact events on the time axis. The impact response template sequence is a long sequence formed by periodically arranging and splicing multiple segments of single fault impact decaying oscillation waveform sequences according to the impact repetition period. This sequence contains multiple impact event waveforms with the same shape, and the time interval between adjacent impact events is strictly equal to the impact repetition period. In this embodiment, the specific execution process of cyclic copying and splicing is as follows: Obtain the value of the impact repetition period, divide the impact repetition period by the sampling time interval of the physical quantity monitoring signal to obtain the impact repetition interval step size in units of sampling points. Obtain the convolution kernel length of the first convolutional layer, divide the convolution kernel length by the impact repetition interval step size and round up to obtain the number of single impact waveforms to be spliced. Create a blank numerical sequence with a length equal to the convolution kernel length as a container for the impact response template sequence. Copy all sampling points of the first segment of single fault impact decaying oscillation waveform sequence sequentially to the starting position of the impact response template sequence. Starting from the end of the first waveform segment, a certain number of blank sampling points are left. The number of blank sampling points is equal to the difference between the larger value of the impact repetition interval step size minus the length of the single-fault impact decay oscillation waveform sequence and zero. This ensures that the beginning of the next waveform segment maintains a time interval exactly equal to the impact repetition period between it and the impact event of the previous waveform segment. After this blank interval, the next single-fault impact decay oscillation waveform sequence is appended. This process of leaving intervals and appending waveforms is repeated until the impact response template sequence container is filled. If there are still empty spaces in the container after appending the last waveform segment, zero-padding is performed at the end as in step S240. Through this cyclic copying and splicing, the impact response template sequence exhibits periodic impact repetition characteristics. The onset, spike, and decay patterns of each impact event remain consistent, and the interval between adjacent impact events is fixed as the impact repetition period.

[0052] Step S240: Align the total length of the shock response template sequence with the length of the convolution kernel of the first convolutional layer. When the total length of the shock response template sequence exceeds the length of the convolution kernel, truncate it. When the total length of the shock response template sequence is less than the length of the convolution kernel, pad it with zeros at the end to obtain the length-aligned shock response template sequence.

[0053] The kernel length of the first convolutional layer is an integer parameter determined during the network architecture design phase, defining the number of sampling points covered by each convolutional kernel in the time dimension. In this embodiment, the specific operation of the alignment process is as follows: obtain the total number of sampling points currently contained in the impulse response template sequence and compare it numerically with the kernel length. If the total length of the impulse response template sequence is greater than the kernel length, starting from the end of the impulse response template sequence, discard the excess sampling points, and only retain a prefix subsequence whose length is exactly equal to the kernel length from the beginning of the sequence. The retained subsequence is the result of the truncation process. If the total length of the impulse response template sequence is less than the kernel length, append zero sampling points equal to the kernel length minus the current total length to the end of the impulse response template sequence. The length of the zero-padding sequence reaches the kernel length. This zero-padding process ensures that the dimension of the impulse response template sequence is numerically equal to the kernel dimension of the first convolutional layer. If the total length of the impulse response template sequence is exactly equal to the kernel length, no additional operation is performed, and it is directly used as the aligned sequence. Alignment processing enables precise positional weight assignment mapping between the shock response template sequence and the convolution kernel.

[0054] Step S250: Map each sequence element in the length-aligned shock response template sequence to the corresponding weight in the convolution kernel of the first convolutional layer, and replace the current value of the corresponding weight with the magnitude of the sequence element to complete the initialization of the weight of the convolution kernel of the first convolutional layer.

[0055] One-to-one mapping refers to the operation of copying the values ​​of one sequence to the same index position of another sequence of the same dimension, according to a one-to-one correspondence between position indices. In this embodiment, during mapping assignment, the length-aligned shock response template sequence is obtained, which contains index positions from zero to the kernel length minus 1 and their corresponding sequence element amplitudes. The kernel weight vector to be initialized in the first convolutional layer is obtained, which can currently store random initialization values ​​or residual values ​​from previous training. Each index position of the shock response template sequence is traversed, the sequence element amplitude at that index position is read, and the amplitude is written to the same index position in the kernel weight vector, replacing the original weight value at that position. If the first convolutional layer contains multiple kernels, the amplitudes of the shock response template sequence are mapped to the corresponding index positions of each kernel weight vector in the same way, so that all kernels share the same set of physically driven weight values ​​in the initial state. After this mapping assignment operation is completed, the weights of all convolutional kernels in the first convolutional layer no longer contain any traces of random initialization, but instead carry the shock response template waveform information that precisely matches the shock dynamics of the equipment failure.

[0056] Step S260: During the sliding convolution process of the physical quantity monitoring signal in the first convolutional layer, the assigned convolutional kernel performs impact response matching with the local segment of the physical quantity monitoring signal at each sliding window position. When the local segment contains an impact pattern similar to the impact response template sequence, a matching response is generated, and a fault-sensitive feature map reflecting the physical guidance of the periodic occurrence location of the fault impact is output.

[0057] Sliding convolution is the forward propagation operation of a convolutional neural network. Specifically, the convolution kernel moves stepwise along the time axis of the input signal with a fixed stride. At each new window position, the kernel weight vector is multiplied element-wise with the corresponding elements of the signal segment covered by the window. The results of all multiplications are then summed to obtain a scalar response value, which represents the degree of matching between the convolution kernel and the signal segment at the current window position. Impulse response matching refers to the fact that when the waveform of the signal segment covered by the sliding window is highly correlated with the waveform of the impulse response template sequence carried by the convolution kernel, the element-wise multiplication and summation operation produces a response output with a large amplitude. When the waveform of the signal segment is uncorrelated with the impulse response template sequence, the response output is at a lower level. In this embodiment, after a physical quantity monitoring signal is fed into the first convolutional layer, the physically initialized convolution kernel performs the above cross-correlation operation at each sliding window position. If the current waveform of a local signal segment happens to be a stable, low-amplitude background vibration before a fault impact, its waveform matches the pre-impact stable segment in the impact response template sequence, and the response value obtained by convolution is in a low amplitude range. If the local signal segment happens to cover a fault impact event, a transient pattern of rapid rising edge followed by decaying oscillation appears in the local signal segment. This pattern highly matches the initiation and decay segments of the impact response template sequence in the convolution kernel, and the response value obtained by convolution suddenly shows a significant positive spike. The response values ​​corresponding to all sliding window positions are arranged in chronological order of the time window, forming a one-dimensional response sequence with a length equal to the length of the input signal after convolution. Multiple convolution kernels independently perform the above sliding convolution process, each outputting its own one-dimensional response sequence. These sequences are stacked along the channel direction, collectively forming a physically guided fault-sensitive feature map. On the time axis of this feature map, the position of the response peak precisely corresponds to the moment when the actual fault impact event occurs in the physical quantity monitoring signal. The spacing between each peak reflects the fault impact repetition period, thus realizing the transformation of the physical quantity monitoring signal into a feature representation with the impact position as the prominent feature under physical constraints.

[0058] Step S300: Input the physically guided fault-sensitive feature map into the physical pooling layer. The physical pooling layer determines the non-uniform downsampling rate according to the preset fault feature frequency ratio, and uses the non-uniform downsampling rate to perform frequency-adaptive feature compression processing on the physically guided fault-sensitive feature map to generate a fault feature representation with reduced dimensionality.

[0059] The physical pooling layer is an intermediate processing layer in the anomaly feature extraction network embedded with physical mechanisms, located after the first convolutional layer and before the classification output layer. Its function is to selectively compress and reduce the temporal dimension of the physically guided fault-sensitive feature map without significantly losing fault impact discrimination information. The preset fault feature frequency ratio refers to the normalized ratio of the fault feature frequency finally retained in step S2216 to the sampling frequency of the physical quantity monitoring signal. The physical meaning of this ratio is the average number of sampling points contained in a complete fault impact repetition cycle. The non-uniform downsampling rate is a downsampling ratio parameter that dynamically changes with the location of the fault impact on the time axis. This parameter maintains a low downsampling ratio at the moment of the fault impact event and in the adjacent period to compress or even retain the original feature elements to a small extent, thereby maintaining the temporal resolution capability of the impact transient details. In the stable running period between two impact events, a higher downsampling ratio is used to compress and merge the feature elements in this period to a larger extent, thereby achieving an effective reduction in the temporal dimension of the overall feature map. Frequency-adaptive feature compression processing refers to assigning differentiated compression intensities to different time regions of the physically guided fault-sensitive feature map based on the frequency information of repeated fault impacts. This ensures that the compressed feature representation maintains high fidelity in the critical impact region and achieves a high compression rate in the impact-independent region. The dimensionality-reduced fault feature representation is a two-dimensional feature map output by the physical pooling layer. Its channel depth is consistent with the input physically guided fault-sensitive feature map, but its length in the time direction is significantly shorter than that of the input feature map. The extent of the shortening depends on the overall average level of the non-uniform downsampling rate. In this embodiment, after receiving the physically guided fault-sensitive feature map, the physical pooling layer first calculates the expected time interval of the fault impact on the time axis based on the preset fault feature frequency ratio. This expected time interval is equal to the reciprocal of the fault feature frequency ratio multiplied by the sampling time interval. Based on this expected time interval, the feature map is divided into two types of intervals along the time axis: alternating impact-related segments and impact-independent segments. The impact-related segments include the time window covered by a preset protection range extending to the left and right from the expected impact occurrence time, while the impact-independent segments are the remaining time regions between two adjacent impact-related segments. For each impact-related segment, a retention-intensive sampling process is performed. This process identifies key feature points such as the impact initiation point, impact termination point, and impact peak location, prioritizing the retention of feature elements within these key points and their neighborhoods, while selectively skipping transitional feature elements between key points that contribute little to the impact morphology. For each impact-independent segment, a compression sampling process is performed. This process merges multiple temporally adjacent feature elements within the segment into a single representative value using an aggregation function, replacing the merged feature element sequence with this representative value, thus achieving a significant dimensionality reduction.The feature fragments of the impact-related segment after retaining dense sampling and the dimensionality-reduced feature fragments of the impact-independent segment after compressive sampling are spliced ​​and combined according to the order of each fragment on the original time axis, while maintaining the independence of each channel in the channel direction. Finally, the fragments are stacked to generate a fault feature representation with reduced dimensionality.

[0060] In one implementation, step S300 may specifically include the following steps S310 to S360: Step S310: Obtain the feature strips of the fault-sensitive feature map of physical guidance along the time dimension. Each feature position in the feature strip corresponds to a sampling moment of the physical quantity monitoring signal on the time axis. The channel direction of the feature strip corresponds to different convolution kernels of the first convolutional layer.

[0061] The physically guided fault-sensitive feature map is a two-dimensional data structure. Its first dimension is the time direction, with a length equal to the output length after the length of the input physical quantity monitoring signal is processed by the sliding convolution of the first convolutional layer. The second dimension is the channel direction, with a depth equal to the total number of convolutional kernels in the first convolutional layer. A feature strip is a one-dimensional feature vector extracted along the time direction from a fixed channel. Each element in this vector is a feature value, and each feature value is associated with three attributes: the numerical value of the feature value (response intensity), the time position index of the feature value in the original feature map, and the channel affiliation label of the feature value. In this embodiment, the feature strip is obtained by traversing each channel index of the physically guided fault-sensitive feature map. For each channel index, all feature values ​​from the start time position to the end time position of that channel are extracted. These feature values ​​are arranged in ascending order of time position index into a one-dimensional array, which is one feature strip. After traversal, a set of feature strips equal to the number of channels is obtained. Each feature strip independently participates in subsequent segment compression processing. During the processing, the feature strips maintain synchronization in time segment division, but operate independently in feature value extraction and aggregation to maintain the differentiated impact response pattern information captured by different convolution kernels.

[0062] Step S320: Based on the preset fault characteristic frequency ratio, determine the expected time interval for the occurrence of fault impact, and divide the characteristic strip into multiple impact-related segments and multiple impact-independent segments along the time direction according to the expected time interval. The impact-related segments include the time of impact occurrence and its adjacent time period, and the impact-independent segments are the stable running period between two impacts.

[0063] The expected time interval refers to the average distance between two adjacent fault impact events on the time axis. Its value is equal to the reciprocal of the preset fault characteristic frequency ratio, multiplied by the sampling time interval of the physical quantity monitoring signal, and converted into a time length in units of sampling points. The impact-related segment is the time interval covered by a preset half-window length extending in both the decreasing and increasing time directions, centered on each expected impact occurrence time. This interval completely encompasses the entire process of impact initiation, spike, and attenuation, as well as a small buffer transition area before and after the impact, ensuring that the complete waveform information of the impact event is protected during subsequent compression. The impact-independent segment is the remaining time region on the time axis between two adjacent impact-related segments. The signal characteristics in this region mainly reflect the stable operating state of the equipment during the interval between two impacts, with a high degree of information redundancy.

[0064] For example, starting from the initial time position of the feature strip, a series of equally spaced expected impact time positions are sequentially marked on the time axis with the expected time interval as the periodic step size. The first expected impact time position can be set at half the expected time interval offset from the initial position of the feature strip. Each subsequent expected impact time position is incremented by one expected time interval from the previous one. Around each expected impact time position, a preset half-window length is subtracted from the position to define the starting boundary of the impact-related segment, and the preset half-window length is added to the position to define the ending boundary of the impact-related segment. The time interval defined by these starting and ending boundaries constitutes an impact-related segment. Between two adjacent impact-related segments, the time interval between the ending boundary of the previous impact-related segment and the starting boundary of the next impact-related segment is defined as an impact-independent segment. The interval between the initial position of the feature strip and the starting boundary of the first impact-related segment, and the interval between the ending boundary of the last impact-related segment and the ending position of the feature strip, are also defined as impact-independent segments. The boundary information of all segments is recorded in the segment division table in the form of start and end time indexes. Subsequent compression operations use this segment division table to determine the segment type to which each feature element belongs.

[0065] Step S330: Perform a retention-dense sampling process on the feature elements within the impact-related segment. By selectively retaining adjacent feature elements within the impact-related segment, intermediate transition features that contribute little to the impact morphology are skipped, thus maintaining the fine temporal structure within the impact-related segment.

[0066] In one implementation, step S330 may specifically include the following steps S331 to S336: Step S331: Extract the feature strip fragments corresponding to the impact-related sections from the physically guided fault-sensitive feature map, and expand the feature elements arranged along the time direction in the feature strip fragments into a one-dimensional feature element sequence. Each feature element in the one-dimensional feature element sequence carries its time position index and channel affiliation mark in the original feature strip.

[0067] In this embodiment, the extraction operation is performed based on the segmentation table generated in step S320: Each segment record marked as an impact-related segment in the segmentation table is traversed, and the start and end time indices of that segment are read. All feature elements contained within the index interval are extracted from the feature strip. The extracted feature elements are arranged sequentially according to their time position indices from smallest to largest, forming a one-dimensional feature element sequence. Each feature element in the sequence contains three attribute fields: response intensity value, i.e., the numerical value of the feature element; time position index, i.e., the sampling point number of the feature element in the original feature strip; and channel attribution marker, i.e., the convolutional kernel channel number to which the feature element belongs. This unfolding operation transforms the local segments originally embedded in the two-dimensional feature map into a one-dimensional linear sequence structure that is convenient for element-by-element analysis and processing.

[0068] Step S332: Compare the response intensities of every two adjacent feature elements in the one-dimensional feature element sequence. When the response intensity of the latter feature element increases sharply compared to the former feature element, mark the position of the feature element with the sharp increase as a candidate point for the start of the impact transient. When the response intensity of the latter feature element decreases sharply compared to the former feature element, mark the position of the feature element with the sharp decrease as a candidate point for the end of the impact transient.

[0069] In this embodiment, the sequence of one-dimensional feature elements is scanned sequentially from the second feature element. At each scan position, the response intensity value of the current feature element is subtracted from the response intensity value of the previous feature element to obtain the response intensity difference. If the response intensity difference is positive and the proportion of the difference to the absolute value of the response intensity of the previous feature element exceeds a preset steep increase ratio threshold, then a steep increase turning point is determined to have occurred at the current feature element position, and the time position index of the current feature element is recorded as a candidate point for the start of an impact transient. The method for determining a steep decrease turning point is as follows: the response intensity value of the previous feature element is subtracted from the response intensity value of the current feature element to obtain the response intensity difference. If the difference is positive and the proportion of the difference to the absolute value of the response intensity of the previous feature element exceeds a preset steep decrease ratio threshold, then a steep decrease turning point is determined to have occurred at the previous feature element position, and the time position index of the previous feature element is recorded as a candidate point for the end of an impact transient. By traversing all adjacent feature element pairs in the one-dimensional feature element sequence and performing the above judgment operation, a set of time position indices for all candidate points of impact transient start and end is obtained. The introduction of steep increase and steep decrease thresholds avoids misjudging small amplitude fluctuations caused by background noise as impact start and end events, thus improving the noise resistance of candidate point extraction.

[0070] Step S333: Obtain the impact attenuation duration indication information. The impact attenuation duration indication information is determined by the damping attenuation trend of the industrial equipment, reflecting the duration of a single fault impact from the start of oscillation to attenuation to near zero amplitude. Pair the candidate points for the start of the impact transient with the candidate points for the end of the impact transient according to the time position index, and measure the time span between each pair of candidate points. Retain the candidate point pairs whose time span is consistent with the impact attenuation duration indication information, and remove isolated candidate points whose time span is inconsistent with the impact attenuation duration indication information, to obtain the impact start point and impact end point after duration consistency screening.

[0071] In one implementation, step S333 may specifically include the following steps S3331 to S3336: Step S3331: Collect the time position indices of all candidate points for the start of the impact transient into a start index set, collect the time position indices of all candidate points for the end of the impact transient into an end index set, and for each start index in the start index set, select the end index from the end index set whose time position is after the start index and whose time difference with the start index does not exceed twice the impact decay duration indication information to form a candidate pairing grid.

[0072] The starting index set is a list consisting of the time and position indices of all candidate starting points for the impact transient recorded in step S332. The ending index set is a list consisting of the time and position indices of all candidate ending points for the impact transient recorded in step S332. The candidate pairing grid is a two-dimensional data structure, where rows correspond to elements in the starting index set and columns correspond to elements in the ending index set. A cell at the intersection of a row and a column in the grid is activated if the ending index value of the column is greater than the starting index value of the row, and the difference between the ending index and the starting index does not exceed twice the impact attenuation duration indication information. The starting and ending indices corresponding to the activated cells form a candidate point pair, which participates in the subsequent duration consistency check. The starting and ending index combinations corresponding to the inactive cells are not subject to subsequent pairing checks because the time span is too large, exceeding the upper limit of the possible duration of a single impact.

[0073] Step S3332: Obtain the impact attenuation duration indication information of the industrial equipment, and determine the allowable deviation range of the time span based on the preset allowable deviation of the duration. The deviation range limits the acceptable deviation between the candidate point and the time span and the impact attenuation duration indication information.

[0074] The preset duration tolerance is a positive numerical parameter, with its physical unit being the number of sampling points. This parameter is used to absorb the actual measurement deviation of the impact duration caused by factors such as small fluctuations in rotational speed, quantization errors in sampling frequency, and nonlinearity of damping characteristics. The deviation interval is defined as a closed interval from the impact attenuation duration indication information minus the duration tolerance to the impact attenuation duration indication information plus the duration tolerance. The upper and lower boundaries of this interval define the maximum allowable positive and negative deviations between the candidate point pair and the physical prior value, respectively. The numerical setting of the duration tolerance must balance the rigor of the screening process with the tolerance for real physical fluctuations.

[0075] Step S3333: For each candidate point pair in the candidate pairing grid, extract the time span between the end index and the start index, compare the time span with the impact attenuation duration indication information, and mark the candidate point pair as a time-consistent candidate point pair when the difference between the time span and the impact attenuation duration indication information falls within the deviation range; otherwise, mark the candidate point pair as a time-inconsistent candidate point pair and remove it.

[0076] In this embodiment, the comparison operation is as follows: The time span is obtained by calculating the difference between the end index value and the start index value in the candidate point pair. The absolute value of the difference between this time span and the impact attenuation duration indication information is calculated. This absolute value is compared with the half-width of the deviation interval determined in step S3332, i.e., the duration allowable deviation. If the absolute value is less than or equal to the duration allowable deviation, the time span is determined to be consistent with the impact attenuation duration indication information, and the candidate point pair is marked as a duration-consistent candidate point pair and retained. If the absolute value is greater than the duration allowable deviation, the time span is determined to be inconsistent with the physical prior duration, indicating that the pairing of the start and end candidate points is likely not generated by the same physical impact event, and this candidate point pair is marked as a duration-inconsistent candidate point pair and eliminated. After this traversal comparison, all remaining are duration-consistent candidate point pairs.

[0077] Step S3334: Check the coverage of all candidate point pairs with consistent duration on the time axis. When the time interval between the end index of the previous period and the start index of the next period in two adjacent candidate point pairs with consistent duration is less than the preset interval compactness requirement, it is determined that there is a time sequence conflict between the two candidate point pairs with consistent duration. Extract the feature response intensity at the start index of each candidate point pair with time sequence conflict, retain the candidate point pair with consistent duration that has the largest feature response intensity, and remove the other conflicting candidate point pairs.

[0078] The preset interval compactness requirement is a time interval threshold, which can be set as a small proportion of the impact repetition cycle, such as one-quarter of the impact repetition cycle. The physical basis for this requirement is that two adjacent fault impact events cannot be infinitely close on the time axis; they should maintain at least a minimum distance determined by the impact decay duration and the impact interval. In this embodiment, the conflict detection process is as follows: all candidate point pairs with consistent durations are sorted in ascending order of their starting indices. Adjacent point pairs are traversed sequentially. For the end index of the previous candidate point pair with consistent duration and the starting index of the next candidate point pair with consistent duration, the difference between them is calculated. If this difference is less than the preset interval compactness requirement, it is determined that there is a temporal conflict between the two candidate point pairs with consistent durations, meaning that at least one of the two point pairs is a false impact marker generated by a false detection. In this case, the feature response intensity corresponding to the starting index position of each of the two conflicting point pairs is extracted, and the magnitude of the response intensity values ​​is compared. The candidate point pair with the larger response intensity value is retained, because a larger response intensity usually corresponds to a more credible real impact event; the conflicting candidate point pair with the smaller response intensity is eliminated. If the response intensities of two conflicting point pairs are very close, the overall energy of the feature elements within each point pair can be used for further discrimination.

[0079] Step S3335: Statistically analyze the time interval sequence between adjacent starting indices of the candidate point pairs with consistent durations retained after conflict removal, analyze the fluctuation uniformity of the time interval sequence, and when the fluctuation uniformity indicates that the dispersion trend of the interval exceeds the preset uniformity requirement, tighten the width of the deviation interval and return to re-execute the candidate point pair marking and removal operation until the fluctuation uniformity of the interval sequence meets the preset uniformity requirement.

[0080] The uniformity of fluctuation is a statistical indicator that measures how close the interval values ​​in a time interval sequence are to each other. In this embodiment, the coefficient of variation is used as the metric. The coefficient of variation is calculated by dividing the standard deviation of the time interval sequence by its arithmetic mean. If the coefficient of variation is greater than the upper limit threshold of the preset uniformity requirement, it indicates that the distribution of the currently selected impact start and end point pairs on the time axis is not uniform enough, and there is an abnormal increase in interval due to missed detection or an abnormal decrease in interval due to false detection, and the overall periodicity is not significant. At this time, the system tightens the allowable deviation of duration, that is, it narrows the width of the deviation interval, for example, by reducing the allowable deviation of duration to a certain proportion of the original, so that the standard for duration consistency verification is more stringent. Using the adjusted narrower deviation interval as the new judgment criterion, the system returns to step S3333 and re-executes the duration consistency comparison and elimination operation in the retained candidate point pair set. This cycle of "tightening the deviation interval - re-screening - verifying uniformity" is repeated until the coefficient of variation of the time interval sequence is reduced to within the preset uniformity requirement. When the uniformity meets the requirements, it indicates that the remaining impact start and end point pairs exhibit a stable periodic arrangement on the time axis, which is consistent with the physical expectation under the repetition frequency of the fault impact.

[0081] Step S3336: Determine the starting index of the candidate point pair with consistent duration that meets the uniformity requirement as the impact start point, and determine the corresponding ending index as the impact end point. Output the impact start point and impact end point as the result after duration consistency screening.

[0082] After the above multi-layered screening and uniformity test, each starting index in the remaining candidate point pairs with consistent duration is officially confirmed as an impact start point, marking the precise start time position of a fault impact event in the feature strip; each corresponding ending index is confirmed as an impact end point, marking the decay termination time position of the same impact event. The impact start points and impact end points are output in pairs, forming a set of impact key points, which are used to construct the subsequent dynamic impact response window.

[0083] Step S334: Take the impact start point that has been screened for duration consistency as the window start point and the impact end point that has been screened for duration consistency as the window end point to form a dynamic impact response window. Within each dynamic impact response window, extract the response intensity of each feature element in the window, arrange the response intensity from large to small, and select a preset number of feature elements at the beginning of the arrangement as core impact features. This preset number does not exceed a preset proportion of the total number of feature elements in the dynamic impact response window.

[0084] The dynamic impact response window is a time interval defined by an impact start point as the left boundary and the impact end point paired with that start point as the right boundary. Each window precisely covers the entire process of a complete impact event from the onset to the end of decay. The preset quantity is the upper limit of the number of feature elements that must be forcibly retained as core impact features within a single dynamic impact response window. The preset ratio is the upper limit of the proportion of core impact features to the total number of feature elements within the window. This ratio is used to prevent the selection of too many core features when the total number of feature elements within the window is small, thus reducing the compression effect. In this embodiment, for each dynamic impact response window, the response intensity values ​​of all feature elements contained within the window are collected, and these response intensity values ​​are sorted in descending order to obtain a sorted list. The preset quantity is calculated as the result of multiplying the total number of feature elements within the window by the preset ratio and rounding down. If this calculated result is less than a preset minimum retention number, the minimum retention number is used as the preset quantity. The feature elements ranked at the top of the sorted list are selected and marked as core impact features. These core impact features embody the local feature information that has the strongest response and contributes the most significant contribution to the expression of impact morphology within the impact window. They will be unconditionally preserved in subsequent compression and will not be discarded by extraction or aggregation operations.

[0085] Step S335: Determine the time region in the one-dimensional feature element sequence that is not covered by any dynamic shock response window as the window-to-window stationary segment. Statistically count the frequency of change of the response intensity of the feature elements in the window-to-window stationary segment between adjacent time positions. Determine the activity indicator of the stationary segment based on the frequency of change. When the activity indicator is in a stationary state, use the first extraction step size. When the activity indicator is in an intermittent fluctuation state, use the second extraction step size. The first extraction step size is greater than the second extraction step size.

[0086] The stable interval between windows refers to the remaining time intervals within the impact-related interval, located between various dynamic impact response windows. These intervals are part of the impact-related interval but are not covered by the impact event itself. The overall characteristic response intensity within them is low, but there may be weak pre-impact fluctuations or aftershock fluctuations. The activity indicator is a measure of the frequency of dynamic changes in characteristic response intensity within the stable interval between windows. It is calculated by traversing all adjacent characteristic element pairs within the interval, counting the number of times the response intensity jumps from a low value to a high value or from a high value to a low value after falling below the preset fluctuation threshold, and dividing the total number of transitions by the total number of characteristic elements within the interval to obtain the change frequency density. The frequency density of changes is compared with a preset activity threshold. If the frequency density is lower than the activity threshold, the activity indicator of the stable segment within the window is determined to be in a stable and unchanging state, indicating that the feature values ​​within this segment have almost no fluctuations and the information redundancy is extremely high. If the frequency density is higher than or equal to the activity threshold, it is determined to be in an intermittent fluctuation state, indicating that there are sporadic and weak fluctuations within this segment, which may contain auxiliary information that has some reference value for subsequent classification. The first extraction step size is a large integer, suitable for the stable and unchanging state, meaning that one feature element is extracted and retained every step size, and the rest are discarded, achieving a large compression. The second extraction step size is a small integer, suitable for the intermittent fluctuation state, meaning that feature elements are extracted at a denser interval, compressing relatively gently to retain fluctuation information. The value of the first extraction step size is greater than the second extraction step size.

[0087] Step S336: Using a determined extraction step size, extract the feature elements within the stable segments between windows at equal intervals to generate a compressed stable segment feature sequence. Merge the core impact features within each dynamic impact response window with the remaining feature elements within that window that were not selected as core impact features in chronological order to form a window-retained feature sequence. Concatenate the window-retained feature sequence with the compressed stable segment feature sequence in chronological order to generate the output feature fragment of the impact-related segment after retaining dense sampling.

[0088] The interval extraction process is as follows: starting from the beginning of a stable segment within a window, one feature element is extracted and retained for every extraction step length corresponding to that segment, while skipped intermediate feature elements are discarded. If the extraction step length is the first extraction step length, the extraction interval is large, and the retained feature elements are relatively sparse; if the extraction step length is the second extraction step length, the extraction interval is small, and the retained feature elements are relatively dense. All extracted retained feature elements are arranged in their original chronological order to form a compressed stable segment feature sequence. For each dynamic impact response window, the core impact feature within the window and the remaining feature elements not marked as core impact features are merged and arranged in ascending order of their respective chronological indices to form the window-in-window retained feature sequence. This sequence uses the core impact feature as its main framework while also including other non-core feature information within the window that may still have auxiliary discriminative value. The feature sequences within the windows corresponding to all dynamic impact response windows in the impact-related segment, as well as the compressed stationary segment feature sequences corresponding to all stationary segments between windows, are spliced ​​together according to the chronological order of each sequence on the original time axis. The spliced ​​overall sequence is the output feature segment of the impact-related segment after the retention dense sampling process. Its time length is significantly reduced compared to the original impact-related segment, but the key morphological information such as the onset, peak, and decay of the impact transient is completely preserved.

[0089] Step S340: Perform compressed sampling on the feature elements in the impact-independent segment, merge the temporally continuous feature elements in the impact-independent segment into a single representative value by aggregation, and replace the merged continuous feature elements with the representative value to generate a dimensionality-reduced feature fragment of the impact-independent segment.

[0090] The compressed sampling process employs a segmented aggregation method to significantly compress the dimensionality of impact-independent segments. Aggregation methods can include calculating the arithmetic mean of the response intensities of feature elements within a continuous segment, taking the maximum value of the response intensities within a continuous segment, or taking the median value of the response intensities within a continuous segment. In this embodiment, the arithmetic mean aggregation method can be used. This involves dividing the sequence of feature elements from the start to the end of an impact-independent segment into several continuous segments with a fixed merging window length. Each segment contains the same number of adjacent feature elements. The merging window length is related to the overall target compression ratio; a larger merging window length results in a higher compression ratio. For each segment, the response intensities of all feature elements within that segment are summed, and then divided by the number of feature elements in that segment. The resulting arithmetic mean is the representative value of that segment. This representative value occupies a feature position in the dimensionality-reduced feature segment, replacing all the original feature elements within that segment. After all segments are processed sequentially, the representative values ​​of each segment are arranged according to their chronological order on the time axis, forming a significantly shortened one-dimensional sequence. This sequence is the dimensionality-reduced feature segment of the impact-independent segment. The large amount of redundant and stable background feature information originally contained in the impact-independent segment is aggregated and condensed into a few representative values, achieving efficient data compression while preserving the overall feature trend of the segment.

[0091] Step S350: Combine the feature segments of the impact-related segment after the retention of dense sampling with the dimension-reduced feature segments of the impact-independent segment after the compression sampling, according to their original time sequence, to form a compressed feature strip.

[0092] In this embodiment, the splicing and combining operation, based on the segment type and start / end boundaries recorded in the segment division table generated in step S320, sequentially reads the processed feature fragments corresponding to each segment in ascending order of time index. If the current segment type is an impact-related segment, the output feature fragment generated in step S336 is retrieved; if the current segment type is an impact-independent segment, the dimensionality-reduced feature fragment generated in step S340 is retrieved. These feature fragments are then concatenated sequentially and appended to a new feature sequence. After all the feature fragments corresponding to all segments have been sequentially spliced, the resulting complete feature sequence is the compressed feature strip. The total length of this feature strip in time is equal to the sum of the lengths of each fragment, significantly shorter than the original feature strip. The splicing process strictly maintains the original temporal sequence between segments, and the feature elements in the compressed feature strip can still be mapped back to the corresponding original time interval.

[0093] Step S360: Stack the compressed feature strips corresponding to each channel in the original channel order to generate a fault feature representation with reduced dimensionality. The length of the fault feature representation with reduced dimensionality in the time direction is shorter than that of the physically guided fault sensitivity feature map.

[0094] In this embodiment, after performing the complete processing flow of steps S310 to S350 on the feature strips of each channel, each channel obtains a non-uniformly compressed feature strip. The length of the compressed feature strips in each channel remains strictly consistent in the time direction because the division of each segment is shared and synchronized across different channels. The compressed feature strips are stacked in ascending order of the original channel index, i.e., the feature strip of channel zero is the first row, the feature strip of channel one is the second row, and so on, until all channels are stacked, forming a new two-dimensional feature map. This two-dimensional feature map is the fault feature representation with reduced dimensionality. Its first dimension is the time direction, and its length is equal to the uniform length of the compressed feature strips, which is significantly shorter than the time dimension length of the physically guided fault-sensitive feature map. The second dimension is the channel direction, and its depth is equal to the total number of convolutional kernels in the first convolutional layer, consistent with the input physically guided fault-sensitive feature map. The reduced fault features indicate that fine transient structural information of the impact is preserved in the impact-related sections, while deep compression is achieved in the impact-independent sections. The overall feature dimensionality is effectively controlled, providing compact and highly discriminative feature inputs for the subsequent classification output layer.

[0095] Step S400: Input the reduced-dimensional fault feature representation into the classification output layer. The classification output layer performs classification mapping on the reduced-dimensional fault feature representation to obtain the confidence score of whether the industrial equipment belongs to the normal state or various abnormal states.

[0096] The classification output layer is the final processing layer of the anomaly feature extraction network embedded with physical mechanisms. Its function is to further map the fault feature representation with reduced dimensionality output from the physical pooling layer into probability confidence scores for each preset state category, thereby completing end-to-end inference from the original physical quantity monitoring signal to the determination of the equipment's operating state. Normal state refers to the operating state of industrial equipment when all physical quantity monitoring indicators are within the design allowable range and no identifiable fault impact feature patterns appear in the physical quantity monitoring signals. Various abnormal states refer to the operating states when specific types of faults have occurred within the equipment and can be characterized by physical impact features. The specific abnormal state categories are defined according to the fault modes of the monitored equipment. For example, for rolling bearing equipment, these may include inner raceway spalling, outer raceway spalling, rolling element pitting, and cage breakage. The confidence score is a continuous value between zero and one output by the classification output layer for each preset state category. This value represents the probability that the equipment is in that state category under the current input physical quantity monitoring signal conditions. The sum of the confidence scores for all state categories is 1. In this embodiment, the classification output layer consists of a fully connected mapping layer and a probability transformation module connected in series. The fully connected mapping layer receives the fault feature representation with reduced dimensionality, flattens and rearranges it into a one-dimensional feature vector, and then performs a linear space transformation on this one-dimensional feature vector using a fully connected mapping weight matrix to generate a category mapping vector with a length equal to the total number of state categories. The probability transformation module applies a flexible maximum transfer function to the category mapping vector to perform probability transformation, converting each element of the category mapping vector into a non-negative probability value with a fixed sum, and outputting a probability distribution vector. The probability values ​​corresponding to the normal state category dimension and each abnormal state category dimension are extracted from the probability distribution vector, which are the normal state confidence score and the confidence scores for each type of abnormal state.

[0097] In one implementation, step S400 may specifically include the following steps S410 to S460: Step S410: Flatten the fault feature representation with reduced dimensionality. While keeping the channel dimension unchanged, rearrange the multidimensional feature map along the time dimension into a one-dimensional feature vector in a single direction. The length of the one-dimensional feature vector is equal to the product of the number of channels and the compressed time length.

[0098] The reduced-dimensionality fault features are represented as a two-dimensional array structure, with a shape equal to the length in the time direction multiplied by the depth in the channel direction. The flattening process is as follows: starting from the first time step, each time step is visited sequentially along the increasing time step direction. At each time step, the feature values ​​of all channels at that time step are read sequentially along the increasing channel index direction. These feature values ​​are arranged into a row segment according to channel order, and then the row segments are concatenated end-to-end according to the time step order. The resulting continuous one-dimensional array is the flattened one-dimensional feature vector, whose total length is strictly equal to the product of the length in the time direction and the depth in the channel direction. During the flattening process, the spatial adjacency relationships between feature values ​​are transformed into positional adjacency relationships in a one-dimensional sequence, but the physical information carried by each feature value remains unchanged.

[0099] Step S420: Obtain the set of fully connected mapping weights in the classification output layer. The initial value of the set of fully connected mapping weights is generated based on the prior information of the energy distribution of fault feature frequency of various abnormal states, so that the weight elements in the set of fully connected mapping weights on the input dimension corresponding to the fault feature frequency have a relative size relationship consistent with the prior information.

[0100] The fully connected mapping weight set is the core parameter in the fully connected mapping layer used to perform linear spatial transformations. Mathematically, it is a two-dimensional weight matrix. The number of rows in this matrix equals the length of the flattened one-dimensional feature vector, i.e., the number of input feature dimensions, and the number of columns equals the total number of preset state categories, i.e., the number of output category dimensions. Prior information on fault feature frequency energy distribution is domain knowledge summarized from numerous historical fault diagnosis cases or obtained through equipment dynamics simulation. This prior knowledge describes the distribution and relative strength of vibration energy caused by fault impact at different fault feature frequencies and their harmonics under different fault conditions. For example, for a bearing outer ring spalling fault, its vibration energy is mainly concentrated at the first and second harmonics of the outer ring impact feature frequency, and the energy proportion at the first harmonic is usually higher than that at the second harmonic. For a bearing inner ring spalling fault, its vibration energy is distributed at the inner ring impact feature frequency and its harmonics. Furthermore, since the inner ring rotates with the spindle, the energy distribution may also be affected by the modulation sidebands. In this embodiment, the initial values ​​of the fully connected mapping weight set are generated as follows: Physical prior indication information is parsed to obtain the main fault feature frequencies corresponding to each type of abnormal state and the energy distribution ratio at these frequencies. For the input dimension positions in the one-dimensional feature vector corresponding to a certain fault feature frequency and its harmonics, the weight elements at these dimension positions in the output column vector corresponding to the abnormal state category in the fully connected mapping weight matrix are assigned initial values ​​proportional to the energy distribution ratio; for input dimension positions that are not associated with any fault feature frequency, small random initial values ​​close to zero are assigned. This physically driven weight initialization method enables the fully connected mapping layer to have differentiated sensitivity to different fault frequency components at the beginning of training, and its attention naturally tends to the frequency dimensions that are more physically meaningful for fault discrimination.

[0101] Step S430: Input the one-dimensional feature vector into the fully connected mapping layer. The fully connected mapping weight set performs a linear space transformation on the one-dimensional feature vector, mapping the one-dimensional feature vector to a space with the same dimension as the total number of abnormal state categories, and generating a category mapping vector.

[0102] In one implementation, step S430 may specifically include the following steps S431 to S436: Step S431: Obtain physical prior indication information. The physical prior indication information includes the proportional relationship of energy distribution of fault characteristic frequencies corresponding to various abnormal states. The proportional relationship of energy distribution indicates the relative energy strength of different fault characteristic frequencies under various abnormal states.

[0103] The physical prior indication information is domain knowledge stored in the form of a data table. Each row of the data table corresponds to an abnormal state category, and each column corresponds to a fault characteristic frequency index. The value of each cell in the table is the normalized energy proportion coefficient at the fault characteristic frequency under that abnormal state. This information can be obtained by performing spectral analysis and statistics on historical fault signals, or by calculating the spectral characteristics of the response signal under various fault conditions using a multibody dynamics simulation model.

[0104] Step S432: Generate an initial mapping weight matrix based on the physical prior indication information. The element distribution of the weight column vectors corresponding to each fault characteristic frequency in the initial mapping weight matrix is ​​consistent with the energy distribution ratio, so that the weight matrix has differentiated initial sensitivity to different fault frequency components.

[0105] In this embodiment, a zero-based matrix of shape (number of input feature dimensions multiplied by the total number of state categories) is first initialized as the initial mapping weight matrix. Each row of the physical prior indication information data table, representing each anomalous state category, is traversed to obtain the output column index corresponding to that anomalous state category, as well as the frequencies of each fault feature and their energy distribution ratios in that row. For each fault feature frequency, its corresponding input dimension index in the flattened one-dimensional feature vector is determined. This input dimension index can be inferred from the temporal position and channel mapping relationship of the fault feature frequency in the physical pooling layer's output feature map. In the initial mapping weight matrix, the matrix element position determined by the row containing the input dimension index and the output column index corresponding to the anomalous state category is located, and the weight value at that position is assigned as the energy distribution ratio multiplied by a basic initialization amplitude factor. For input dimensions in the one-dimensional feature vector that do not correspond to any fault feature frequencies, their initial weight values ​​in each output column are set to small, uniformly distributed random numbers centered at zero.

[0106] Step S433: Assign the initial mapping weight matrix to the mapping weight set of the fully connected mapping layer as the weight state at the start of training.

[0107] The mapping weight set is the weight matrix parameter of the fully connected mapping layer. Through an assignment operation, its internal values ​​are directly replaced with the initial mapping weight matrix generated in step S432, while the bias term parameter is initialized to an all-zero vector. After the assignment is completed, the weight distribution of the fully connected mapping layer at the initial training time incorporates the physical prior fault frequency sensitivity bias.

[0108] Step S434: In each weight adjustment cycle of the training process, extract the weight subset corresponding to each fault feature frequency in the mapping weight set, analyze the current distribution status of the elements in the weight subset, and compare it with the energy distribution ratio in the physical prior indication information to obtain the distribution deviation indication.

[0109] The weight adjustment cycle refers to one parameter update iteration during the training process. The analysis of the current distribution is as follows: For each fault feature frequency, extract the weight elements of all output category columns in the corresponding input dimension row of the mapping weight set, and calculate the normalized energy proportion vector of these weight elements. Compare this normalized energy proportion vector with the prior energy distribution proportion vector of the fault feature frequency in each abnormal state category in the physical prior indication information, and calculate the vector distance between the two, for example, using cosine distance or Euclidean distance as a distribution deviation indicator. The distribution deviation indicator quantifies the degree of deviation of the current weight distribution of the fully connected mapping layer from the physical prior distribution.

[0110] Step S435: Generate weight adjustment constraints based on the distribution deviation indication. Apply the weight adjustment constraints and the weight adjustment requirements generated by the classification loss together to the mapping weight set, restricting the update direction of the mapping weight set, so that the difference between the updated weight distribution and the physical prior indication information is reduced.

[0111] In one implementation, step S435 may specifically include the following steps S4351 to S4356: Step S4351: In each weight update cycle of the training process, obtain the element value distribution of the weight subset corresponding to the fault feature frequency in the current mapped weight set, and generate the current sensitivity distribution shape.

[0112] The current sensitivity distribution pattern refers to the normalized energy ratio distribution of the frequencies of each fault feature across all output category dimensions in the current mapping weight set.

[0113] Step S4352: Compare the current sensitivity distribution pattern with the energy distribution ratio in the physical prior indication information to determine the deviation direction and deviation magnitude in each fault characteristic frequency dimension, and form a distribution deviation indication.

[0114] The deviation direction is a sign indicating whether the current weight ratio is too large or too small relative to the prior ratio, and the deviation magnitude is the absolute value of the difference between the two.

[0115] Step S4353: Based on the distribution deviation indication, generate a constraint adjustment signal for the weight dimension with deviation. The direction of the constraint adjustment signal is to make the weight move in a direction that conforms to the physical prior indication information. The magnitude of the constraint adjustment signal is positively correlated with the deviation magnitude.

[0116] For weighted elements with large deviations, the amplitude of the constraint adjustment signal is correspondingly larger to apply a stronger physical regression pull; for weighted elements with small deviations or no deviation, the amplitude of the constraint adjustment signal is correspondingly smaller or zero.

[0117] Step S4354: Obtain the classification adjustment signal generated by backpropagation of the classification loss function. Superimpose the constraint adjustment signal and the classification adjustment signal according to the corresponding element positions to form a composite adjustment signal. When the two signals are in the same direction, they are superimposed and enhanced; when they are in opposite directions, they cancel each other out.

[0118] The classification loss function can be the cross-entropy loss function. The classification adjustment signal is the partial derivative of the cross-entropy loss with respect to each element of the fully connected mapping weight matrix, calculated layer by layer through the backpropagation algorithm. The superposition operation is an element-wise algebraic summation. When the update directions of the classification adjustment signal and the constraint adjustment signal are the same on that element, their superposition generates a stronger update driving force; when their directions are opposite, they cancel each other out, forming a dynamic balance between physical constraints and data-driven forces.

[0119] Step S4355: Update each weight element in the mapping weight set using the composite adjustment signal. The updated weight element value is determined by the current value, the constraint adjustment signal, and the classification adjustment signal.

[0120] The update algorithm adopts the weight update rule of the adaptive moment estimation optimizer, that is, it maintains the first moment estimate and the second moment estimate of each weight element. The first moment estimate is the exponentially weighted moving average of the historical gradient including the composite adjustment signal, and the second moment estimate is the exponentially weighted moving average of the square of the historical gradient. The weight update amount is equal to the learning rate multiplied by the first moment estimate divided by the sum of the square root of the second moment estimate and the minimum constant. The updated weight value is obtained by subtracting the weight update amount from the current weight value.

[0121] Step S4356: Repeat the process of obtaining the distribution pattern, generating deviation indication, generating constraint adjustment signal and updating weights until the difference between the sensitivity distribution pattern of the mapped weight set and the physical prior indication information enters the preset stable range. At this time, stop applying the constraint adjustment signal and only retain the classification adjustment signal for subsequent updates.

[0122] The preset stability interval is a near-zero tolerance range for the distribution deviation indicator. When the deviation indicator falls within this stability interval, it indicates that the sensitivity distribution of the mapping weight set has fully matched the physical prior. After this, the physical constraints have fulfilled their purpose, and the training process is driven solely by the classification loss for fine convergence.

[0123] Step S436: Use the set of mapping weights after weight adjustment constraints to perform a linear transformation on the one-dimensional feature vector to generate a class mapping vector. The magnitude of each element in the class mapping vector is affected by the feature frequency sensitivity under physical prior constraints.

[0124] After multiple rounds of training with physical consistency constraints, the weight distribution of the fully connected mapping layer has fully learned the mapping relationship from the energy distribution of fault feature frequencies to specific state categories within a physically reasonable framework. The category mapping vector generated by this set of weights can more reliably reflect the differentiated contributions of different fault frequency components to the discrimination of each abnormal state.

[0125] Step S440: Perform probability transformation on the category mapping vector, convert each element in the category mapping vector into a non-negative value, and normalize all elements so that the sum of all transformed elements is a fixed value, thus obtaining the probability distribution vector.

[0126] The probability transformation process employs a flexible maximum transfer function. In this embodiment, each element in the category mapping vector is iterated, and its exponential function value is calculated using the base of the natural logarithm to obtain the corresponding exponent value. The exponent values ​​of all elements are summed to obtain the total exponent. For each element, its exponent value is divided by the total exponent, and the quotient is the probability value corresponding to that element. The sum of the probability values ​​of all elements is strictly equal to 1, and each probability value falls within the interval between 0 and 1. The vector composed of these probability values ​​is the probability distribution vector.

[0127] Step S450: Extract the elements corresponding to the normal state category from the probability distribution vector as normal state confidence scores, and extract the elements corresponding to each abnormal state category as abnormal state confidence scores. The confidence scores represent the probability that the industrial equipment is in the corresponding state category.

[0128] In this embodiment, a fixed index mapping relationship exists between each dimension of the probability distribution vector and a preset state category. For example, index 0 corresponds to the normal state category, index 1 corresponds to the inner ring peeling state category, index 2 corresponds to the outer ring peeling state category, and index 3 corresponds to the rolling element pitting state category. Based on this mapping relationship, the element value at index 0 in the probability distribution vector is read as the confidence score for the normal state, the element value at index 1 is read as the confidence score for the inner ring peeling state, and so on. Each confidence score comprehensively reflects the degree of confidence of the abnormal feature extraction network embedded with the physical mechanism in determining whether the current input physical quantity monitoring signal belongs to each state category.

[0129] Step S460: Output the confidence score for normal state and the confidence scores for various abnormal states as the classification mapping results of the current operating state of industrial equipment.

[0130] The classification mapping results can be output in the form of structured data, such as a set of key-value pairs containing the status category name and the corresponding confidence score value, which can be provided to the upper-level monitoring system or human-machine interface for subsequent status alarm interpretation and maintenance decision support.

[0131] Step S500: The anomaly feature extraction network embedded with physical mechanism is trained using a loss function that includes a physical consistency regularization term. The physical consistency regularization term is used to constrain the spectral energy distribution of the convolution kernel of the first convolutional layer to be concentrated at the preset fault feature frequency and its harmonics in the frequency domain. During the inference stage, the real-time collected physical quantity monitoring signal is input into the trained anomaly feature extraction network embedded with physical mechanism, and then processed sequentially through the first convolutional layer, the physical pooling layer and the classification output layer to obtain the anomaly state recognition result.

[0132] As one implementation method, step S500 involves training the anomaly feature extraction network embedding physical mechanisms using a loss function that includes a physical consistency regularization term. Specifically, this may include the following steps S510-S560: Step S510: Obtain a training signal set, which contains physical quantity monitoring signal samples labeled with state category labels. The state category labels indicate whether the sample belongs to a normal state or a specific abnormal state type.

[0133] The training signal set is a database of physical quantity monitoring signals accumulated over a long period during the equipment's historical operation or collected after various typical faults were artificially introduced through a fault simulation test bench. Each sample is a fixed-length signal segment extracted from long-term continuous monitoring signals. The extraction method can be overlapping extraction using a sliding window or extraction triggered by the equipment's operating conditions. Each sample is associated with a state category label, which is stored in integer code form. For example, label value zero represents a normal state, label value one represents an inner race fault state, label value two represents an outer race fault state, and label value three represents a rolling element fault state. The training signal set undergoes preprocessing before training. Preprocessing operations include removing DC bias components from the signals and normalizing the signal amplitude according to the standard deviation to ensure that the signals under different operating conditions have comparable amplitude ranges.

[0134] Step S520: Input each sample in the training signal set into the anomaly feature extraction network that embeds the physical mechanism, and generate the prediction confidence score set corresponding to the sample through forward propagation of the first convolutional layer, physical pooling layer and classification output layer.

[0135] The predicted confidence score set is consistent with the output format of step S460, containing the network's confidence scores for the current input sample belonging to normal and various abnormal states. The forward propagation process proceeds sequentially as follows: the first convolutional layer performs sliding convolution on the input sample to generate a physically guided fault-sensitive feature map; the physical pooling layer performs non-uniform downsampling compression on the feature map to generate a dimensionality-reduced fault feature representation; and the classification output layer flattens, performs fully connected mapping, and performs probability transformation on the compressed features to generate the predicted confidence score set.

[0136] Step S530: Use the classification loss metric function to compare the predicted confidence score set with the state category label of the sample to generate a classification loss term. The classification loss term reflects the deviation between the prediction result and the true state category.

[0137] The classification loss metric uses the cross-entropy loss function. In this embodiment, the cross-entropy loss term is calculated as follows: For each training sample, the confidence score corresponding to its true state category label is extracted from the set of predicted confidence scores for that sample. The natural logarithm of this confidence score is calculated and multiplied by negative one to obtain the individual cross-entropy loss for that sample. The arithmetic mean of the individual cross-entropy losses of all training samples in the current batch is then calculated, and the resulting average is the classification loss term for that batch.

[0138] Step S540: Perform time-domain to frequency-domain mapping on the current weight sequence of the convolution kernel of the first convolutional layer to obtain the frequency domain representation of the convolution kernel. The frequency domain representation includes each frequency position and its corresponding spectral intensity.

[0139] The mapping from the time domain to the frequency domain can be performed using the Fast Fourier Transform (FFT) algorithm. In this embodiment, the mapping process is performed on the weight sequence of each convolutional kernel in the first convolutional layer at the current training iteration. The FFT transforms the one-dimensional weight sequence in the time domain into a one-dimensional complex sequence in the frequency domain, with the length of the frequency domain sequence being the same as the length of the time domain sequence. Each index position in the frequency domain sequence corresponds to a digital frequency, which can be converted into an actual frequency in Hertz. The conversion relationship is that the digital frequency equals the actual frequency divided by the sampling frequency of the physical quantity monitoring signal and then multiplied by the number of FFT points. The spectral intensity is taken as the modulus of each complex element in the frequency domain complex sequence, that is, the square root of the sum of the squares of the real part and the squares of the imaginary part. Each convolutional kernel yields a spectral intensity sequence, and the spectral intensity sequences of all convolutional kernels together constitute the frequency domain representation of the convolutional kernel.

[0140] In one implementation, step S540 may specifically include the following steps S541 to S546: Step S541: Extract all convolutional kernel weights at the current time step of the first convolutional layer to form a set of convolutional kernel weight sequences. Each sequence in the set of convolutional kernel weight sequences corresponds to the weight distribution of a convolutional kernel in the time domain.

[0141] In this embodiment, the weight vectors of all convolutional kernels at the current iteration time are read from the model parameter storage of the first convolutional layer. The length of each weight vector is the length of the convolutional kernel, and the weight vectors of all convolutional kernels constitute a set of convolutional kernel weight sequences.

[0142] Step S542: Perform time-domain to frequency-domain mapping processing on each convolution kernel weight sequence in sequence, convert the time-domain weight sequence into a frequency-domain complex representation, and take the amplitude of each frequency component in the frequency-domain complex representation as the spectral intensity to generate the spectral intensity sequence corresponding to the convolution kernel.

[0143] The specific steps of the Fast Fourier Transform (FFT) algorithm for mapping processing are as follows: The convolution kernel weight sequence is padded with zeros to a power of two to facilitate the FFT operation. A butterfly recursive decomposition is performed on the padded sequence, dividing it into two groups based on index parity (odd or even), performing Fourier transforms on each group separately, and then recombining them. This recursion continues until the sequence length reaches two, at which point a two-point Discrete Fourier Transform (DFT) is directly performed. The result is a complex array. The modulus of each complex number in this array is taken as the spectral intensity, yielding the spectral intensity sequence of the convolution kernel.

[0144] Step S543: Merge the spectral intensity sequences corresponding to each convolutional kernel point by point according to the same frequency position to obtain the average spectral intensity sequence. Each element in the average spectral intensity sequence represents the overall spectral response intensity of all convolutional kernels in the first convolutional layer at that frequency position.

[0145] For each frequency position index, iterate through the spectral intensity sequence of all convolution kernels, read the spectral intensity value at that frequency position, sum the values ​​of all convolution kernels at that position, divide by the total number of convolution kernels, and the resulting average value is the element value of the average spectral intensity sequence at that frequency position.

[0146] Step S544: In the average spectral intensity sequence, determine the frequency index position corresponding to the preset fault characteristic frequency, and extract the spectral intensity of the frequency index position and the adjacent micro-intervals as the fault fundamental frequency spectral intensity.

[0147] The frequency resolution is obtained by dividing the sampling frequency of the physical quantity monitoring signal by the number of points in the Fast Fourier Transform. The preset fault characteristic frequency is then divided by the frequency resolution, and the quotient is rounded to the nearest integer to obtain the frequency index position corresponding to the fault characteristic frequency. Adjacent small intervals are defined as ranges offset to the left and right of this frequency index position by two frequency resolution indices to encompass any slight drift that the fault characteristic frequency may experience during actual operation. The maximum or average spectral intensity at all frequency positions within this interval is extracted as the fault fundamental frequency spectral intensity.

[0148] Step S545: Based on the preset frequency multiplier, determine the set of frequency multiplier index positions corresponding to the multiple frequencies of the fault characteristic frequency, extract the spectral intensity at each position in the set of frequency multiplier index positions, and merge it with the spectral intensity of the fault fundamental frequency to form a set of spectral intensity of the fault frequency band.

[0149] The preset harmonics can be the first, second, and third harmonics of the fault characteristic frequency. For each harmonic, the harmonic value is calculated as the fault characteristic frequency multiplied by the harmonic. The harmonic value is then divided by the frequency resolution and rounded to the nearest integer to obtain the harmonic index position. For each harmonic index position and the small intervals between the two frequency resolution indices to its left and right, the maximum value of the spectral intensity within each interval is extracted as the spectral intensity corresponding to that harmonic. The spectral intensity of the fault fundamental frequency is then combined with the spectral intensities of all harmonics to form the spectral intensity set of the fault frequency band.

[0150] Step S546: Extract the spectral intensity of the remaining frequency positions in the average spectral intensity sequence, excluding the fault characteristic frequency index position and its overtone index position, into a set of non-fault frequency band spectral intensity. Generate a spectral energy concentration indicator based on the set of fault frequency band spectral intensity and the set of non-fault frequency band spectral intensity.

[0151] In one implementation, after step S546, the following steps S547 to S5412 may also be included: Step S547: Generate a spectrum adjustment signal acting on the convolution kernel of the first convolutional layer using the spectrum energy concentration indicator. The direction of the spectrum adjustment signal is to concentrate the spectrum energy distribution of the convolution kernel towards the fault characteristic frequency and its harmonics.

[0152] For example, an ideal filter template is designed in the frequency domain. This template has a unity gain at the fault characteristic frequency index and its harmonics, and a zero gain at all other frequencies. Subtracting the average spectral intensity sequence from the ideal filter template point by point yields a spectral bias sequence. This spectral bias sequence reflects the difference between the current frequency domain distribution of the convolution kernel and the ideal physical distribution. Transforming the spectral bias sequence back from the frequency domain to the time domain using an inverse fast Fourier transform yields a time-domain spectral adjustment signal. This signal directs the convolution kernel's time-domain weights towards eliminating the spectral bias, thus further concentrating the kernel's spectral energy at the fault characteristic frequency and its harmonics.

[0153] Step S548: Obtain the classification adjustment signal generated by the backpropagation of the classification loss term, and perform directional coordination processing on the spectral adjustment signal and the classification adjustment signal in the convolution kernel weight space to generate a composite adjustment signal. When there is a conflict between the spectral adjustment signal and the classification adjustment signal in the weight update direction, the adjustment component that is consistent with the frequency sensitive direction of the fault feature is retained first.

[0154] For each element position of the convolution kernel weights, obtain the values ​​of the spectral adjustment signal and the classification adjustment signal at that position, and calculate their signs. If the signs are the same, the composite adjustment signal is the weighted sum of the spectral adjustment signal and the classification adjustment signal, with the weighting coefficient reflecting the trade-off between physical consistency constraints and classification accuracy requirements. If the signs are opposite, it indicates that the classification loss aims to adjust the weight element in a direction away from the physical concentration distribution. In this case, the composite adjustment signal discards the component of the classification adjustment signal and only takes the component of the spectral adjustment signal at that element position, or takes the component of the classification adjustment signal projected onto that element in the same direction as the spectral adjustment signal. This ensures that the convolution kernel weights will not weaken their established physical consistency no matter how they are updated.

[0155] Step S549: Apply the composite adjustment signal element by element to the weights of each convolution kernel in the first convolutional layer, update the convolution kernel weights, and the updated convolution kernel weights reflect both the classification accuracy requirement and the spectral energy concentration requirement.

[0156] For example, the new value applied as the convolution kernel weight is equal to the current value minus the learning rate multiplied by the value of the composite adjustment signal on that element.

[0157] Step S5410: Perform time-domain to frequency-domain mapping processing on the updated convolution kernel weights again, and redetermine the set of spectrum intensity of faulty frequency bands and the set of spectrum intensity of non-faulty frequency bands to generate an updated spectrum energy concentration indicator.

[0158] This step repeats steps S541 to S546 to obtain an indication of the spectral energy concentration of the updated convolution kernel.

[0159] Step S5411: Compare the updated spectrum energy concentration indicator with the previous spectrum energy concentration indicator. If the updated spectrum energy concentration indicator shows that the trend of spectrum energy gathering towards the fault characteristic frequency and its harmonics is enhanced, then maintain the current convolution kernel weights and proceed to the next training iteration.

[0160] The criterion for determining an enhanced clustering trend is that the updated spectral energy concentration indicator value is greater than the value before the update.

[0161] Step S5412: If the updated spectrum energy concentration indicator shows that the spectrum energy concentration trend has not increased, repeat the steps of generating spectrum adjustment signal, direction coordination processing, updating convolution kernel weights and redetermining spectrum energy concentration indicator until the spectrum energy concentration indicator shows that the concentration trend has increased, and then proceed to the next training iteration.

[0162] This loop ensures that each adjustment to the convolution kernel weights actually promotes the concentration of spectral energy toward the fault characteristic frequency and its harmonics, avoiding a regression in physical consistency.

[0163] Step S550: In the frequency domain representation, determine the first frequency position corresponding to the preset fault characteristic frequency and the second frequency position set corresponding to the harmonic of the fault characteristic frequency, extract the spectral intensity at the first frequency position and the spectral intensity at each position in the second frequency position set to generate a fault frequency band energy indicator, and extract the spectral intensity at all frequency positions other than the first frequency position and the second frequency position set to generate a non-fault frequency band energy indicator.

[0164] The first frequency position is the frequency index position corresponding to the fault characteristic frequency determined in step S544. The second frequency position set is the set consisting of all the harmonic index positions determined in step S545. The fault band energy indication is the sum of the spectral intensity value at the first frequency position and the spectral intensity values ​​at all frequency positions in the second frequency position set. The non-fault band energy indication is the difference obtained by subtracting the fault band energy indication from the total spectral intensity of all frequency positions in the average spectral intensity sequence.

[0165] Step S560: Based on the comparison between the energy indicators of the fault band and the energy indicators of the non-fault band, generate a physical consistency regularization term, merge the classification loss term and the physical consistency regularization term to obtain a joint loss, and use the joint loss to adjust the internal connection strength of the abnormal feature extraction network embedded with the physical mechanism until the preset training stopping condition is met.

[0166] The comparison relationship is handled as follows: the energy indicator of the non-faulty frequency band is divided by the sum of the energy indicator of the faulty frequency band and a very small constant. The resulting ratio is then multiplied by a positive physical consistency weight coefficient. The product is the physical consistency regularization term. The physical consistency weight coefficient is a hyperparameter that controls the strength of physical constraints. The larger its value, the stronger the constraint on the physical consistency of the convolutional kernel during training. The joint loss is equal to the classification loss term generated in step S530 plus the physical consistency regularization term. After obtaining the joint loss, the partial derivatives of the joint loss with respect to the anomalous feature extraction network embedding the physical mechanism are calculated by automatic differentiation. This includes the convolutional kernel weights and biases of the first convolutional layer, and the mapping weight set and bias terms of the fully connected mapping layer. The adaptive moment estimation optimizer is used to update each parameter based on these partial derivatives. The preset training stopping conditions can be that the joint loss does not decrease in a specified number of consecutive iterations, or the total number of training iterations reaches the preset training cycle limit, or the classification accuracy reaches the target performance level on the independent validation set. Once the stopping condition is met, the training process terminates, and the connection strength parameters of each layer of the network are fixed as final values ​​and stored as the trained model file. During the inference phase, this model file is loaded into the online monitoring system. Real-time collected physical quantity monitoring signals are input into the network in a streaming manner, sequentially passing through the physical driving feature extraction of the first convolutional layer, the frequency adaptive compression of the physical pooling layer, and the state mapping of the classification output layer. Finally, the abnormal state identification result is output, providing a decision-making basis for predictive maintenance and fault tracing in intelligent manufacturing production lines.

[0167] This invention provides a computer-readable storage medium storing computer-executable instructions or a computer program. When the computer-executable instructions or the computer program are executed by a processor, the processor will execute the artificial intelligence-based intelligent manufacturing anomaly state identification method provided in this invention. For example, ... Figure 3 The method for identifying abnormal states in intelligent manufacturing based on artificial intelligence is shown.

[0168] In some embodiments, the computer-readable storage medium may be a read-only memory (ROM), random access memory (RAM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory, magnetic surface memory, optical disk, or CD-ROM, etc.; or it may be a device that includes one or any combination of the above-mentioned memories.

[0169] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for identifying abnormal states in intelligent manufacturing based on artificial intelligence, characterized in that, The method includes: Acquire physical quantity monitoring signals collected by industrial equipment during operation; An anomaly feature extraction network embedded with physical mechanisms is obtained. The anomaly feature extraction network embedded with physical mechanisms includes a first convolutional layer, a physical pooling layer, and a classification output layer connected in sequence. The first convolutional layer is physically initialized. The initial values ​​of the convolutional kernel weights of the first convolutional layer are assigned using the impact response function corresponding to the dynamic behavior of equipment faults. This makes the forward propagation process of the first convolutional layer equivalent to performing a fault feature transformation under physical constraints on the physical quantity monitoring signal, generating a physically guided fault-sensitive feature map. The physically guided fault-sensitive feature map is input into the physical pooling layer. The physical pooling layer determines a non-uniform downsampling rate based on a preset fault feature frequency ratio, and uses the non-uniform downsampling rate to perform frequency-adaptive feature compression processing on the physically guided fault-sensitive feature map to generate a fault feature representation with reduced dimensionality. The reduced-dimensional fault feature representation is input into the classification output layer, which performs classification mapping on the reduced-dimensional fault feature representation to obtain a confidence score indicating whether the industrial equipment belongs to a normal state or various abnormal states. The abnormal feature extraction network embedded with the physical mechanism is trained using a loss function that includes a physical consistency regularization term. The physical consistency regularization term is used to constrain the spectral energy distribution of the convolution kernel of the first convolutional layer in the frequency domain to be concentrated at the preset fault feature frequency and its harmonics. During the inference phase, the real-time collected physical quantity monitoring signal is input into the trained abnormal feature extraction network embedded with the physical mechanism, and then processed sequentially through the first convolutional layer, the physical pooling layer, and the classification output layer to obtain the abnormal state identification result.

2. The method for identifying abnormal states in intelligent manufacturing based on artificial intelligence according to claim 1, characterized in that, The physical initialization process for the first convolutional layer involves assigning initial values ​​to the kernel weights of the first convolutional layer using an impact response function corresponding to the equipment fault dynamics. This makes the forward propagation process of the first convolutional layer equivalent to performing a fault feature transformation under physical constraints on the physical quantity monitoring signal, generating a physically guided fault-sensitive feature map, including: The inherent structural indication information of the industrial equipment is analyzed, and the operating frequency indication information and geometric dimension indication information that affect the dynamic response mode of the fault impact are separated from it. The operating frequency indication information represents the periodicity of the rotation of the main shaft of the equipment, and the geometric dimension indication information represents the diameter of the rolling element distribution circle, the individual diameter of the rolling element, the total number of rolling elements, and the contact tilt angle. Using the operating frequency indication information and the geometric dimension indication information, the fault impact response generation process is driven to generate a damped oscillation waveform sequence simulating a single fault impact. The damped oscillation waveform sequence is composed of the rapid rising edge of the impact oscillation stage and the exponential envelope falling edge of the damping decay stage. Obtain the fault impact repetition frequency generated by the operating frequency indication information and the geometric dimension indication information, and cyclically copy and splice the damped oscillation waveform sequence according to the impact repetition period determined by the fault impact repetition frequency. Maintain an interval consistent with the impact repetition period between the end of the previous waveform sequence and the beginning of the next waveform sequence to form an impact response template sequence with continuous impact interval characteristics. The total length of the shock response template sequence is aligned with the length of the convolution kernel of the first convolutional layer. When the total length of the shock response template sequence exceeds the length of the convolution kernel, it is truncated. When the total length of the shock response template sequence is less than the length of the convolution kernel, it is padded with zeros at the end to obtain the length-aligned shock response template sequence. Each sequence element in the length-aligned shock response template sequence is mapped one by one to the corresponding position weight in the convolution kernel of the first convolutional layer, and the magnitude of the sequence element is used to replace the current value of the corresponding position weight to complete the initialization assignment of the convolution kernel weight of the first convolutional layer. During the sliding convolution process of the physical quantity monitoring signal in the first convolutional layer, the assigned convolutional kernel performs impact response matching with local segments of the physical quantity monitoring signal at each sliding window position. When a local segment contains an impact pattern similar to the impact response template sequence, a matching response is generated, and a fault-sensitive feature map reflecting the physical guidance of the periodic occurrence location of the fault impact is output.

3. The method for identifying abnormal states in intelligent manufacturing based on artificial intelligence according to claim 2, characterized in that, The process of using the operating frequency indication information and the geometric dimension indication information to drive the fault impact response generation process, generating a damped oscillation waveform sequence simulating a single fault impact, includes: The repetition frequency indication information of the fault impact event is generated from the operating frequency indication information and the geometric dimension indication information, and the repetition frequency indication information is used as the fault characteristic frequency, which represents the number of times the fault impact occurs per unit time; the impact response natural frequency indication information of the industrial equipment is obtained, which reflects the resonance frequency of the equipment structure under fault impact excitation. Using the natural frequency of the impact response as the oscillation frequency, a sinusoidal oscillation fundamental wave with the same periodicity as the natural frequency of the impact response is generated. The sinusoidal oscillation fundamental wave is a continuous periodic waveform with stable amplitude, and its period duration is the reciprocal of the natural frequency of the impact response. Obtain damping attenuation trend indication information reflecting the damping characteristics of the industrial equipment structure. The damping attenuation trend indication information is determined by the material internal friction and the friction characteristics of the joint surface. Use the damping attenuation trend indication information to generate an exponential attenuation envelope. The exponential attenuation envelope shows a decreasing trend in the time domain, gradually approaching zero from the initial time to subsequent time. The instantaneous amplitude of the sinusoidal oscillation fundamental wave and the corresponding amplitude of the exponentially decaying envelope are modulated point by point, so that the oscillation amplitude of the sinusoidal oscillation fundamental wave changes synchronously according to the decay trend of the exponentially decaying envelope, thereby generating a decaying oscillation waveform of a single impact. The position of the transient response peak at the moment of impact onset is determined in the damped oscillation waveform, and a pre-impact stable segment is intercepted forward from the position of the transient response peak. The waveform amplitude in the pre-impact stable segment is stable relative to the amplitude at the position of the transient response peak. The pre-impact stable segment is spliced ​​together with the damped oscillation waveform, with the tail end of the pre-impact stable segment connected to the start end of the damped oscillation waveform, to form a complete single-fault impact damped oscillation waveform sequence that includes the stable state before impact and the damped oscillation state after impact.

4. The method for identifying abnormal states in intelligent manufacturing based on artificial intelligence according to claim 3, characterized in that, The generation of repetition frequency indication information for the fault impact event from the operating frequency indication information and the geometric dimension indication information includes: The rolling element pitch circle diameter indication information, rolling element diameter indication information, rolling element quantity indication information, and contact angle indication information are extracted from the geometric dimension indication information. The rolling element pitch circle diameter indication information is the diameter of the center distribution circle of the rolling element, the rolling element diameter indication information is the diameter of a single rolling element, and the contact angle indication information is the angle between the contact direction of the rolling element and the raceway and the radial plane. Based on the rolling element pitch circle diameter indication information, rolling element diameter indication information, and contact angle indication information, and combined with the operating frequency indication information, the inner ring impact characteristic frequency of the rolling element passing through the inner ring fault point, the outer ring impact characteristic frequency of the rolling element passing through the outer ring fault point, and the rolling element impact characteristic frequency caused by the rolling element's own defects are generated respectively. The inner ring impact characteristic frequency, the outer ring impact characteristic frequency and the rolling element impact characteristic frequency are aggregated into a fault impact event frequency set, and the fault impact event frequency set contains at least one type of fault corresponding to the repetition frequency. Obtain the fault type indication information recorded in the historical fault records of the industrial equipment, and select a specific fault impact frequency that is consistent with the fault type indication information from the fault impact event frequency set as the preferred repetition frequency. When the historical fault records are missing, each candidate impact frequency in the fault impact event frequency set is extracted in sequence, driving subsequent steps to generate their respective corresponding fault sensitive feature maps, and obtaining the classification mapping response intensity of the classification output layer to each fault sensitive feature map, and determining the retention order of the candidate impact frequencies based on the response intensity. The final retained fault impact frequency is output as the repetition frequency information of the fault impact event, and the repetition frequency information is used to determine the splicing period of the subsequent impact response template sequence.

5. The method for identifying abnormal states in intelligent manufacturing based on artificial intelligence according to claim 1, characterized in that, The step involves inputting the physically guided fault-sensitive feature map into the physical pooling layer. The physical pooling layer determines a non-uniform downsampling rate based on a preset fault feature frequency ratio and uses the non-uniform downsampling rate to perform frequency-adaptive feature compression processing on the physically guided fault-sensitive feature map to generate a fault feature representation with reduced dimensionality. This includes: The fault-sensitive feature map of the physical guidance is obtained by expanding the feature strips along the time dimension. Each feature position in the feature strip corresponds to a sampling time of the physical quantity monitoring signal on the time axis, and the channel direction of the feature strip corresponds to different convolution kernels of the first convolutional layer. Based on the preset fault characteristic frequency ratio, the expected time interval for the occurrence of fault impact is determined, and the characteristic strip is divided into multiple impact-related segments and multiple impact-independent segments along the time direction according to the expected time interval. The impact-related segments include the time of impact occurrence and its adjacent time period, and the impact-independent segments are the stable running period between two impacts. The feature elements within the impact-related segment are subjected to a retention-dense sampling process. By selectively retaining adjacent feature elements within the impact-related segment, intermediate transition features that contribute little to the impact morphology are skipped, thus maintaining the fine temporal structure within the impact-related segment. Compressed sampling is performed on the feature elements within the impact-independent segment. Temporally continuous feature elements within the impact-independent segment are merged into a single representative value by aggregation. The merged continuous feature elements are replaced by the representative value to generate a dimensionality-reduced feature fragment of the impact-independent segment. The feature fragments of the impact-related segment after retaining dense sampling and the dimensionality-reduced feature fragments of the impact-independent segment after compressed sampling are spliced ​​together in the original time sequence to form a compressed feature strip. The compressed feature strips corresponding to each channel are stacked in the original channel order to generate a fault feature representation with reduced dimensionality. The length of the fault feature representation in the time direction is shorter than that of the physically guided fault sensitivity feature map.

6. The method for identifying abnormal states in intelligent manufacturing based on artificial intelligence according to claim 5, characterized in that, The process of performing dense sampling on the feature elements within the impact-related segment includes: Extract the feature strip fragments corresponding to the impact-related sections from the physical-guided fault-sensitive feature map, and expand the feature elements arranged along the time direction in the feature strip fragments into a one-dimensional feature element sequence. Each feature element in the one-dimensional feature element sequence carries its time position index and channel affiliation mark in the original feature strip. The response intensities of every two adjacent feature elements in the one-dimensional feature element sequence are compared. When the response intensity of the latter feature element increases sharply compared to the former feature element, the position of the feature element with the sharp increase is marked as a candidate point for the start of the impact transient. When the response intensity of the latter feature element decreases sharply compared to the former feature element, the position of the feature element with the sharp decrease is marked as a candidate point for the end of the impact transient. Obtain the impact attenuation duration indication information, pair the impact transient start candidate point and the impact transient end candidate point according to the time position index, measure the time span between each pair of candidate points, retain candidate point pairs whose time span is consistent with the impact attenuation duration indication information, and remove isolated candidate points whose time span is inconsistent with the impact attenuation duration indication information to obtain the impact start point and impact end point after duration consistency screening; The impact start point that has been screened for consistency of duration is taken as the start point of the window, and the impact end point that has been screened for consistency of duration is taken as the end point of the window, forming a dynamic impact response window. Within each dynamic impact response window, the response intensity of each feature element in the window is extracted, the response intensity is arranged from large to small, and a preset number of feature elements at the beginning of the arrangement are selected as core impact features. The preset number does not exceed a preset proportion of the total number of feature elements in the dynamic impact response window. The time region in the one-dimensional feature element sequence that is not covered by any dynamic shock response window is defined as the window-to-window stable segment. The frequency of change of the response intensity of the feature elements in the window-to-window stable segment between adjacent time positions is counted. The activity indicator of the stable segment is determined based on the frequency of change. When the activity indicator is in a stable and constant state, a first extraction step size is used. When the activity indicator is in an intermittent fluctuation state, a second extraction step size is used. The first extraction step size is larger than the second extraction step size. Using a determined extraction step size, feature elements within the stable segments between windows are extracted at equal intervals to generate a compressed stable segment feature sequence. The core impact features within each dynamic impact response window are merged with the remaining feature elements within that window that were not selected as core impact features in chronological order to form a window-retained feature sequence. The window-retained feature sequence is then concatenated with the compressed stable segment feature sequence in chronological order to generate the output feature fragment of the impact-related segment after retention-intensive sampling processing.

7. The method for identifying abnormal states in intelligent manufacturing based on artificial intelligence according to claim 6, characterized in that, The process of obtaining the impact attenuation duration indication information involves pairing the impact transient start point candidate point with the impact transient end point candidate point according to their time position index, measuring the time span between each pair of candidate points, retaining candidate point pairs whose time span matches the impact attenuation duration indication information, and removing isolated candidate points whose time span does not match the impact attenuation duration indication information. This yields the impact start point and impact end point after duration consistency screening, including: The time position indices of all candidate points for the start of the impact transient are collected into a start index set, and the time position indices of all candidate points for the end of the impact transient are collected into an end index set. For each start index in the start index set, the end index whose time position is after the start index and whose time difference with the start index does not exceed twice the impact attenuation duration indication information is selected from the end index set to form a candidate pairing grid. The impact attenuation duration indication information of the industrial equipment is obtained, and the allowable deviation range of the time span is determined according to the preset allowable deviation of the duration. The deviation range limits the acceptable deviation between the candidate point and the time span and the impact attenuation duration indication information. For each candidate point pair in the candidate pairing grid, the time span between the end index and the start index is extracted. The time span is compared with the impact attenuation duration indication information. When the difference between the time span and the impact attenuation duration indication information falls within the deviation range, the candidate point pair is marked as a candidate point pair with consistent duration. Otherwise, the candidate point pair is marked as a candidate point pair with inconsistent duration and is removed. Check the coverage of all candidate point pairs with consistent duration on the time axis. When the time interval between the end index of the previous period and the start index of the next period in two adjacent candidate point pairs with consistent duration is less than the preset interval compactness requirement, it is determined that there is a time sequence conflict between the two candidate point pairs with consistent duration. Extract the feature response intensity at the start index of each candidate point pair with time sequence conflict, retain the candidate point pair with the largest feature response intensity, and remove the other conflicting candidate point pairs. The time interval sequence between adjacent starting indices of the candidate point pairs with consistent durations retained after conflict removal is statistically analyzed. The fluctuation uniformity of the time interval sequence is analyzed. When the fluctuation uniformity indicates that the dispersion trend of the interval exceeds the preset uniformity requirement, the width of the deviation interval is tightened and adjusted, and the candidate point pair marking and removal operation is re-executed until the fluctuation uniformity of the interval sequence meets the preset uniformity requirement. The starting index of the candidate point pair with consistent duration that meets the uniformity requirement is determined as the impact start point, and the corresponding ending index is determined as the impact end point. The impact start point and the impact end point are output as the result after duration consistency screening.

8. The method for identifying abnormal states in intelligent manufacturing based on artificial intelligence according to claim 1, characterized in that, The reduced-dimensionality fault feature representation is input into the classification output layer, which performs classification mapping on the reduced-dimensionality fault feature representation to obtain a confidence score indicating whether the industrial equipment belongs to a normal state or various abnormal states, including: The reduced-dimensional fault feature representation is flattened, and the multidimensional feature map is rearranged along the time dimension into a one-dimensional feature vector in a single direction while keeping the channel dimension unchanged. The length of the one-dimensional feature vector is equal to the product of the number of channels and the compressed time length. Obtain the set of fully connected mapping weights in the classification output layer. The initial value of the set of fully connected mapping weights is generated based on the prior information of the energy distribution of fault feature frequency of various abnormal states, so that the weight elements in the set of fully connected mapping weights on the input dimension corresponding to the fault feature frequency have a relative size relationship consistent with the prior information. The one-dimensional feature vector is input into the fully connected mapping layer, and the one-dimensional feature vector is transformed by the set of fully connected mapping weights to map the one-dimensional feature vector to a space with the same dimension as the total number of abnormal state categories, thereby generating a category mapping vector. The category mapping vector is subjected to probability transformation processing, in which each element in the category mapping vector is converted into a non-negative value, and all elements are normalized so that the sum of all transformed elements is a fixed value, thus obtaining a probability distribution vector. The elements corresponding to the normal state category in the probability distribution vector are extracted as normal state confidence scores, and the elements corresponding to each abnormal state category are extracted as abnormal state confidence scores. The confidence scores represent the probability that the industrial equipment is in the corresponding state category. The confidence scores for the normal state and the confidence scores for each type of abnormal state are output as the classification mapping results of the current operating state of the industrial equipment.

9. The method for identifying abnormal states in intelligent manufacturing based on artificial intelligence according to claim 8, characterized in that, The step of inputting the one-dimensional feature vector into the fully connected mapping layer, and performing a linear space transformation on the one-dimensional feature vector by the fully connected mapping weight set, includes: Obtain physical prior indication information, which includes the proportional relationship of energy distribution of fault characteristic frequencies corresponding to various abnormal states, and the proportional relationship of energy distribution indicates the relative energy strength of different fault characteristic frequencies under various abnormal states. An initial mapping weight matrix is ​​generated based on the physical prior indication information. The element distribution of the weight column vector corresponding to each fault characteristic frequency in the initial mapping weight matrix is ​​consistent with the energy distribution ratio, so that the weight matrix has differentiated initial sensitivity to different fault frequency components. The initial mapping weight matrix is ​​assigned to the mapping weight set of the fully connected mapping layer as the weight state at the start of training. In each weight adjustment cycle of the training process, the weight subset corresponding to each fault feature frequency in the mapping weight set is extracted, the current distribution status of the elements in the weight subset is analyzed, and it is compared with the energy distribution ratio in the physical prior indication information to obtain the distribution deviation indication. Based on the distribution deviation indication, a weight adjustment constraint is generated. The weight adjustment constraint and the weight adjustment requirement generated by the classification loss are applied together to the mapping weight set to restrict the update direction of the mapping weight set, thereby reducing the difference between the updated weight distribution and the physical prior indication information. The one-dimensional feature vector is linearly transformed using the set of mapping weights after weight adjustment constraints to generate a category mapping vector. The magnitude of each element in the category mapping vector is affected by the feature frequency sensitivity under physical prior constraints.

10. A computer system, characterized in that, include: Memory is used to store executable instructions or computer programs. The processor, when executing computer-executable instructions or computer programs stored in the memory, implements the artificial intelligence-based intelligent manufacturing abnormal state identification method according to any one of claims 1 to 9.

Images