An industrial robot fault diagnosis method based on segment context enhancement and logical decoupling

Abstract

The present application relates to a kind of industrial robot fault diagnosis methods based on segmented context enhancement and logic decoupling, comprising: collecting the synchronous time series data in the process of six-axis industrial robot operation, synchronous time series data is deconstructed into joint level sub-sequence;Dynamic injection semantic anchor operator between adjacent joint sub-sequence in joint level sub-sequence, and after being spliced with constraint operator, obtain enhanced input stream, inject thought chain guide instruction at the end of enhanced input stream, obtain complete structured prompt text input stream;Complete structured prompt text input stream is used to stage training to large language model, generates optimal large language model, for the fault diagnosis of six-axis industrial robot.The present application does not change the premise of large model bottom parameter, by structured prompt word strategy, the keeping capacity of model to long-range complex fault rule is significantly improved, ensure the diagnosis precision of single and composite fault under small sample condition.

Description

Technical Field

[0001] This invention relates to the field of industrial diagnostics technology, and in particular to a fault diagnosis method for industrial robots based on segmented context enhancement and logic decoupling. Background Technology

[0002] In the evolution of Industry 4.0, industrial robots have become the core pillar of automated production lines, undertaking key tasks such as precision assembly, continuous trajectory welding, and heavy-duty handling.

[0003] Taking a six-axis robot as an example, this type of robot drives its joints (Joint, J, ...) through six independent servo motors. to The joints work together, and each joint contains a reducer (Reducer, R). to ) and motor (Motor, M, to This forms a highly complex nonlinear dynamic system. To ensure its reliable operation, a series of high-frequency sensors are integrated within the system to monitor current, torque, angular velocity, and vibration signals in real time. The time-series data generated by these sensors are like the robot's "vital signs," containing rich information about its operational status. However, electromagnetic interference in the industrial environment, dynamic coupling between multiple joints, and dynamic changes in the workload cause these signals to exhibit strong non-stationarity and noise interference, posing significant technical challenges to the early identification of faults.

[0004] In the fault system of a six-axis robot, diagnostic logic typically needs to handle two types of scenarios. The first is a single fault, where the anomaly is limited to a specific physical joint or transmission unit. For example, wear on the reducer might cause an abnormal shift in the current curve of joint 3, and the mapping relationship is usually a simple point-to-point pattern. The second is a more complex compound fault, which is jointly induced by anomalies in the coordinated movement of two or more joints, reflecting logical disorder at the system level. For example, when the reducer of joint 1 and the motor of joint 2 simultaneously exhibit dynamic response hysteresis, the system classifies it as a specific type of compound functional failure. In small-sample environments, due to the lack of sufficient labeled fault data, how to accurately extract key features from these high-dimensional, long-time-series signals and achieve logical alignment is a current research focus in the field of industrial diagnostics.

[0005] Furthermore, existing mainstream solutions such as the FD-LLM framework and ChatTS framework typically employ end-to-end cue engineering, which involves constructing long text cue words to input diagnostic rules and all sensor data into the model at once. That is, a continuous text is used to input the background setting, mapping rules, and all joint time series data into the model. However, this approach has revealed significant performance bottlenecks in practical applications. 1. Long-series signals generated by a six-axis robot, after being converted into text tokens, often exceed the effective coverage of large model attention mechanisms, leading to the phenomenon of "long-distance rule forgetting." When the model reads the end of the sequence, the response strength to the complex composite fault mapping rules defined in the initial segment will significantly decrease.

[0006] 2. Continuous and high-frequency pure numerical streams lack semantic anchors, making large models prone to getting caught in local numerical fluctuations when calculating self-attention weights. This leads to problems in large models when processing data such as... arrive When dealing with cross-joint features, it is impossible to maintain continuous attention to subtle abnormal features of specific joints, which can easily lead to logical illusions or random misjudgments. Summary of the Invention

[0007] To address the problems of long-sequence failures and logical illusions in the diagnosis of six-axis industrial robots, the present invention aims to provide an industrial robot fault diagnosis method based on segmented context enhancement and logical decoupling. The method aims to significantly improve the model's ability to maintain long-range complex fault rules through a structured cue word strategy without changing the underlying parameters of the large model, thus ensuring the diagnostic accuracy of single and compound faults under small sample conditions.

[0008] To achieve the above objectives, the present invention provides the following solution: A fault diagnosis method for industrial robots based on segmented context enhancement and logic decoupling includes: Synchronous timing data during the operation of a six-axis industrial robot is collected, and the synchronous timing data is deconstructed into joint-level subsequences. Semantic anchoring operators are dynamically injected between adjacent joint subsequences in the joint-level subsequence, and then concatenated with constraint operators to obtain an enhanced input stream. At the end of the enhanced input stream, a thought chain guidance instruction is injected to obtain a complete structured prompt text input stream. The large language model is trained in stages using the complete structured prompt text input stream to generate the optimal large language model, which is then used for fault diagnosis of the six-axis industrial robot.

[0009] Optionally, the constraint operator includes: ; in, Define the semantics of the expert and define the behavioral guidelines of the model as an expert in precision diagnosis of industrial robots in the current context. This is a pre-defined set of fault mapping relationships, which includes a pre-defined function that maps the state of underlying physical components to system-level fault categories. The fault mapping relationship set includes: the first fault type corresponding to the fault of the third joint gear reducer, the second fault type corresponding to the fault of the second joint motor, the third fault type corresponding to the fault of the fourth joint gear reducer, the fourth fault type corresponding to the simultaneous fault of the first joint gear reducer and the second joint motor, the fifth fault type corresponding to the simultaneous fault of the first and third joint gear reducers, and the sixth fault type corresponding to the simultaneous fault of the third and fourth joint gear reducers.

[0010] Optionally, obtaining the enhanced input stream includes: ; in, To enhance the input stream, For text concatenation operators, As dynamic semantic anchors, during the processing of enhanced input streams in a large language model, periodic attention reset actions are performed to reiterate the currently investigated joint indices and force the model to activate the pre-defined set of fault mapping relationships in the constraint operators. For the first Joint subsequences, For constraint operators.

[0011] Optionally, obtaining the complete structured prompt text input stream includes: ; in, For a complete structured prompt text input stream, To enhance the input stream, For text concatenation operators, For dynamic semantic anchors, For the first Joint subsequences, For constraint operators, Provides guidance instructions for the thought process.

[0012] Optionally, staged training of the large language model using the complete structured prompt text input stream includes: The first stage of training involves using the large language model to call the underlying physical component states of the constraint operators in the complete structured prompt text to perform feature judgment on the fluctuation pattern of each joint sub-sequence and obtain a set of candidate abnormal components. The feature judgment includes: low-frequency heavy-load wear characteristics of the reducer or high-frequency electromagnetic abnormal characteristics of the motor. The second stage of training involves logically aligning the candidate abnormal component set with the fault mapping relationship set of the constraint operators in the complete structured prompt text based on the large language model. If the candidate abnormal component set matches the fault type in the fault mapping relationship set, a system-level fault category label or a normal label is output. If the candidate abnormal component set does not match the fault type in the fault mapping relationship set, a normal label is output.

[0013] Optionally, obtaining the set of candidate abnormal components includes: ; in, For the set of candidate exception components, This indicates a decelerator malfunction in joints 1-6. This indicates a motor malfunction in joints 1-6.

[0014] Optionally, the output of the system-level fault category label includes: ; in, This is a system-level fault category label or a normal label.

[0015] The beneficial effects of this invention are as follows: This invention alleviates the performance degradation problem of large models when processing long industrial signals through an attention-awakening mechanism, enabling the attention mechanism to maintain sensitivity to diagnostic rules over long distances. Compared with traditional end-to-end direct output methods, the decoupled design of this invention provides a clear inference chain for the model, significantly enhancing the interpretability of diagnostic results.

[0016] In the context of multivariable collaborative work in six-axis robots, this invention effectively suppresses random illusions caused by numerical interference in the model, improves the accuracy of composite fault identification to industrial-grade application requirements, and possesses strong versatility, requiring only adjustment of constraint operators. The rules defined in the code can be quickly migrated to different robot platforms. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an industrial robot fault diagnosis method based on segmented context enhancement and logic decoupling, according to an embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] like Figure 1 As shown, this embodiment discloses an industrial robot fault diagnosis method based on segmented context enhancement and logic decoupling, including: Collect synchronous timing data during the operation of a six-axis industrial robot, and deconstruct the synchronous timing data into joint-level subsequences; Semantic anchoring operators are dynamically injected between adjacent joint subsequences in the joint-level subsequence, and then concatenated with constraint operators to obtain an enhanced input stream. At the end of the enhanced input stream, a thought chain guidance instruction is injected to obtain a complete structured prompt text input stream. A large language model is trained in stages using a complete structured prompt text input stream to generate an optimal large language model for fault diagnosis of a six-axis industrial robot.

[0022] Specifically, this embodiment constructs a hierarchical diagnostic architecture of "benchmark pre-setting - feature deconstruction - semantic anchoring - logical decoupling" to transform complex multidimensional industrial signals into reasoning tasks that can be understood and logically aligned by a large language model. The specific implementation steps are as follows: Step 1: Diagnostic Criteria and Logical Priors Construction: Before the inference task begins, the guiding constraint operator is first constructed. This is used to provide the model with the necessary industry expert background and underlying logical constraints. The mathematical expression of this operator is: ; in, Define the semantics of the expert, define the behavioral guidelines of the model as an "industrial robot precision diagnostic expert" in the current context, and clarify its tasks; This is a predefined set of fault mapping relationships, which predefines the correspondence functions from the state of underlying physical components to system-level fault categories. For example, define... (Single Failure) and (Complex fault).

[0023] Step 2: Physical spatial partitioning and mapping of time-series data: To accommodate the attention window limitations of the model when processing high-dimensional data, this step performs spatial segmentation of the original monitoring sequence according to the physical dimension. Let... This invention deconstructs the complete set of sensor values ​​synchronously acquired by a six-axis robot according to their physical meaning into... indivual( Joint-level subsequences: ; in, For the first Feature sequence fragments of each joint ( ), which serves as a local input for the model to recognize the state of the underlying components.

[0024] Step 3: Inter-segment semantic anchoring and attention reawakening: To address the issue that long sequences of numerical input can easily cause the model to forget preceding rules, this method uses adjacent joint subsequences... and Dynamic injection of semantic anchoring operators The semantic anchoring operator is a semantic tuple dynamically generated based on the physical topology of the current input sequence, and its mathematical construct is expressed as: ; in, This represents the physical dimension spatial anchoring variable, used to identify the currently scanned first dimension. Each joint subsequence is used to establish a strong semantic mapping relationship between the current numerical segment and the corresponding joint entity; This indicates an attention reset instruction, used to create semantic intervals between adjacent numerical subsequences, alleviate attention decay caused by continuous digital streams, and guide the model to refocus on previous fault mapping rules and abnormal features of the current joint.

[0025] Guiding constraint operator With joint subsequence and semantic anchoring operators Concatenate to form an enhanced input stream Its expression is: ; in, This is a text concatenation operator that linearly connects instructions, data, and anchors in a temporal order.

[0026] Step 4: Two-stage logical decoupling reasoning based on the thought chain: Building Enhanced Data Streams Afterwards, End-injection thought chain guidance instructions Construct a complete structured prompt text input stream that is ultimately input into a large language model. : ; Mind Chain Guiding Instructions The guided model follows a two-stage reasoning path rule of "component identification first, then logical alignment".

[0027] The model is based on The guided execution of the decoupled two-stage diagnostic logic: Phase A, Abnormal Joint Assembly Extraction: The model applies to each joint subsequence Perform a deep feature scan. Although the input is fused temporal data at the joint level, the model needs to combine... Based on pre-existing physical knowledge, the model analyzes whether the fluctuation pattern of the joint signal is more consistent with the low-frequency heavy-load wear characteristics of the reducer or the high-frequency electromagnetic anomaly characteristics of the motor. The model then outputs a specific set of candidate anomaly components. The elements of this collection are taken from the virtual split component library: ; in, Representing the The joint's decelerator is malfunctioning. Representing the The joint's motor is malfunctioning. This step enables in-depth analysis from "joint-level timing" to "component-level entity." For example, the model output is: This means that the reducer of the second joint and the motor of the third joint are determined to have abnormal characteristics.

[0028] Phase B, Logical Mapping Function Execution: In determining the set of abnormal components Then, the model is called. Pre-defined combinational logic functions Perform logical alignment. This stage no longer focuses on underlying numerical fluctuations, but only performs symbolic Boolean logic checks: ; in, This is the final output's system-level fault category label or normal label. The components included match specific patterns in the rule base M (such as " Fault and If a fault is detected (e.g., "Fault 4"), the corresponding system-level fault category label is output (e.g., "Fault 4"). If no combinational logic is hit, "Normal" is output according to the settings. This step separates complex feature recognition from simple symbol matching in the inference time domain.

[0029] To further illustrate the technical implementation process of this invention, this embodiment takes a six-axis industrial robot fault diagnosis scenario as an example to demonstrate how this invention utilizes segmented context enhancement and logic decoupling mechanisms to perform structured processing on long-term industrial monitoring data and complete the entire process from model preparation to test and diagnostic output. This embodiment simulates the operational diagnostic process of a six-axis industrial robot under normal, single fault, and combined fault conditions.

[0030] Initial Input and Diagnostic Scenario: The diagnostic object is multi-dimensional timing monitoring data collected during the operation of a six-axis industrial robot. The motor feedback current is used as the main input signal, and the sampling frequency is once per second. Each input sample contains synchronous timing data of joints 1 to 6 within the same operating cycle.

[0031] In this embodiment, the system needs to determine whether the robot state corresponding to the input data is in a normal state or one of the following fault types: Fault 1: Joint 3 gear reducer malfunction; Fault 2: Joint 2 motor malfunction; Fault 3: Joint 4 gear reducer malfunction; Fault 4: Both the gear reducer in joint 1 and the motor in joint 2 fail simultaneously; Fault 5: Both joint 1 and joint 3 gear reducers fail simultaneously; Fault 6: Both joint 3 and joint 4 gear reducers fail simultaneously.

[0032] The system takes a set of timing monitoring sequences of a six-axis robot to be diagnosed as input and outputs the corresponding system-level fault category label or normal status label.

[0033] Phase 1: Model Preparation and Rule Construction The system first selects a pre-trained large language model with general reasoning capabilities as the base model and constructs diagnostic benchmarks and logical priors for it. At this stage, no fine-tuning of the underlying parameters of the base model is performed; instead, its diagnostic capabilities are enhanced through input-side prompt structure design. The system first defines the model's role as an industrial robot precision diagnostic expert in the prompts, clarifying its task to identify abnormal components and output fault categories based on the input joint timing data. Subsequently, the system establishes a fault mapping relationship set, establishing a correspondence between underlying abnormal components and system-level fault labels. For example, "Joint 3 reducer abnormal" is mapped to "Fault 1," and "Joint 1 reducer abnormal and Joint 2 motor abnormal" is mapped to "Fault 4." This rule base serves as the basis for subsequent logical mapping decisions.

[0034] Phase Two: Training Sample Construction and Segmented Augmentation Processing The system performs structured preprocessing on the training samples. First, each original monitoring sequence is segmented according to physical joints, resulting in multiple joint-level subsequences corresponding to joints 1 to 6. Each subsequence retains only the local temporal features of the corresponding joint, which reduces the attention burden when directly inputting long sequences.

[0035] After obtaining the joint-level subsequences, the system injects semantic anchoring information between adjacent joint-level subsequences. This semantic anchoring information is used to repeatedly remind the currently troubleshooting joint and the rules for continuous fault monitoring, thereby periodically waking the model's memory of previous rules when processing continuous numerical streams. This approach can alleviate the problems of rule forgetting and attention drift under long-sequence input.

[0036] After segmentation and semantic enhancement, the system concatenates the diagnostic baseline, joint-level subsequences, and semantic anchoring content into a structured enhanced input stream, which is used for prompt template optimization and diagnostic process verification in the subsequent training phase.

[0037] Phase Three: Two-Stage Diagnostic Training Guided by the Thinking Chain After the structured enhanced input stream is constructed, the system further injects a thought chain guidance instruction at the end of the input to clarify that the model should perform diagnosis in the order of "first identifying abnormal components, then mapping logical rules".

[0038] In this stage, the system uses training set samples to iteratively optimize the prompt template, enabling the model to gradually and stably follow a two-stage logical decoupling path. In the first stage, the model identifies abnormal components based on the temporal variation characteristics of each joint-level subsequence, outputting a set of candidate abnormal components, such as identifying an abnormality in the motor at joint 2 or an abnormality in the reducer at joint 3. In the second stage, based on the identification results from the first stage, the model calls a pre-set set of fault mapping relationships to perform logical alignment, converting the abnormal component combinations into system-level fault categories.

[0039] Through repeated validation with training samples, the system continuously adjusts the rule expressions, semantic anchor positions, and thought chain instructions in the prompt template, enabling the model to form stable diagnostic output behavior without requiring parameter updates.

[0040] This embodiment uses the following structured template: Diagnostic Criteria and Logic Priors: You are a six-axis robot diagnostic expert. Determine whether the input data corresponds to a normal or faulty robot state. Collect feedback current from the robot motors at a sampling frequency of once per second. Strictly adhere to the rules for using fault labels. Define the rules: Normal: Normal; Fault 1: Joint 3 gear reducer malfunction; Fault 2: Joint 2 motor malfunction; Fault 3: Joint 4 gear reducer malfunction; Fault 4: Both the gear reducer in joint 1 and the motor in joint 2 fail simultaneously; Fault 5: Both joint 1 and joint 3 gear reducers fail simultaneously; Fault 6: Both joint 3 and joint 4 gear reducers fail simultaneously.

[0041] Multidimensional feature structured injection: Semantic anchoring of joint 1: Currently scanning the feature sequence of joint 1. Please pay attention to the abnormal features of each axis and continuously anchor the fault logic mapping rules. Semantic anchoring of joint 2: Currently scanning the feature sequence of joint 2. Please pay attention to the abnormal features of each axis and continuously anchor the fault logic mapping rules. Continue until the semantic anchoring of joint 6, then output the result; Logical decoupling reasoning: Review the mapping rules again: Fault 1 (R3), Fault 2 (M2)...

[0042] Please follow these two steps: Step 1: Analyze the above data to identify which components are abnormal. Step 2: Based on the rules, output the final fault type.

[0043] In one possible implementation, an example of fault diagnosis during the testing phase is given: During the testing phase, a monitoring sequence of a six-axis robot that was not trained was input into the system. The system first performs joint-level segmentation and semantic anchoring enhancement according to predetermined rules, and then simultaneously inputs the enhanced structured input stream and thought chain guidance instructions into the large language model.

[0044] For example, in a test sample, the model first analyzes the subsequences of joints 1 to 6 segment by segment. In the first stage of identification, the model determines that the reducer corresponding to joint 1 has abnormal features, and at the same time identifies that the motor corresponding to joint 2 has abnormal features. Therefore, the output set of candidate abnormal components is "reducer abnormal in joint 1, motor abnormal in joint 2".

[0045] Subsequently, during the second-stage logical mapping process, the system performs logical matching based on a pre-set rule base. Since the candidate abnormal component set matches the combined rule "simultaneous failure of joint 1 gear reducer and joint 2 motor", the system ultimately outputs the diagnostic result as "fault 4".

[0046] If the model does not identify any abnormal components in another test sample, or the identification result does not match any preset fault combination rules, the system outputs a "normal" status.

[0047] As can be seen from the above implementation examples, this invention does not rely on fine-tuning the parameters of a large language model. Instead, it achieves stable fault diagnosis under long-term industrial data through joint-level segmentation, inter-segment semantic anchoring, thought chain guidance, and two-stage logical decoupling.

[0048] Compared to the end-to-end approach that inputs all rules and the original long sequence into the model at once, this embodiment can more effectively maintain the model's continuous focus on fault rules, reducing logical illusions and random misjudgments caused by purely numerical input. Furthermore, because the model outputs the abnormal component identification results first and then performs rule mapping, the diagnostic process has stronger interpretability, making it particularly suitable for small-sample industrial diagnostic scenarios where both single and complex faults coexist.

[0049] This embodiment solves the "rule amnesia" problem that is common in large-scale language models when processing high-dimensional signals from multiple joints of robots by embedding semantic operators with memory wake-up function between joint sub-sequence segments, and achieves an effective extension of the upper limit of long text processing.

[0050] This embodiment forces the model to separate the two actions of "feature extraction" and "rule classification" in the time domain at the output end, and uses the identification of the preceding anomaly set to provide a clear logical anchor for the subsequent symbol mapping, thereby eliminating the diagnostic illusion caused by the diffusion of attention distribution.

[0051] This technical solution, which achieves highly reliable diagnosis without relying on model parameter fine-tuning and only through input-side logic reconstruction, has extremely high technical protection value for small-sample fault identification in distributed industrial control systems.

[0052] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A fault diagnosis method for industrial robots based on segmented context enhancement and logic decoupling, characterized in that, include: Synchronous timing data during the operation of a six-axis industrial robot is collected, and the synchronous timing data is deconstructed into joint-level subsequences. Semantic anchoring operators are dynamically injected between adjacent joint subsequences in the joint-level subsequence, and then concatenated with constraint operators to obtain an enhanced input stream. At the end of the enhanced input stream, a thought chain guidance instruction is injected to obtain a complete structured prompt text input stream. The large language model is trained in stages using the complete structured prompt text input stream to generate the optimal large language model, which is then used for fault diagnosis of the six-axis industrial robot.

2. The industrial robot fault diagnosis method based on segmented context enhancement and logic decoupling according to claim 1, characterized in that, The constraint operators include: ； in, Define the semantics of the expert and define the behavioral guidelines of the model as an expert in precision diagnosis of industrial robots in the current context. This is a pre-defined set of fault mapping relationships, which includes a pre-defined function that maps the state of underlying physical components to system-level fault categories. The fault mapping relationship set includes: the first fault type corresponding to the fault of the third joint gear reducer, the second fault type corresponding to the fault of the second joint motor, the third fault type corresponding to the fault of the fourth joint gear reducer, the fourth fault type corresponding to the simultaneous fault of the first joint gear reducer and the second joint motor, the fifth fault type corresponding to the simultaneous fault of the first and third joint gear reducers, and the sixth fault type corresponding to the simultaneous fault of the third and fourth joint gear reducers.

3. The industrial robot fault diagnosis method based on segmented context enhancement and logic decoupling according to claim 1, characterized in that, Obtaining the enhanced input stream includes: ； in, To enhance the input stream, For text concatenation operators, As dynamic semantic anchors, during the processing of enhanced input streams in a large language model, periodic attention reset actions are performed to reiterate the currently investigated joint indices and force the model to activate the pre-defined set of fault mapping relationships in the constraint operators. For the first Joint subsequences, For constraint operators.

4. The industrial robot fault diagnosis method based on segmented context enhancement and logic decoupling according to claim 1, characterized in that, Obtaining the complete structured prompt text input stream includes: ； in, For a complete structured prompt text input stream, To enhance the input stream, For text concatenation operators, For dynamic semantic anchors, For the first Joint subsequences, For constraint operators, Provides guidance instructions for the thought process.

5. The industrial robot fault diagnosis method based on segmented context enhancement and logic decoupling according to claim 1, characterized in that, Staged training of a large language model using the complete structured prompt text input stream includes: The first stage of training involves using the large language model to call the underlying physical component states of the constraint operators in the complete structured prompt text to perform feature judgment on the fluctuation pattern of each joint sub-sequence and obtain a set of candidate abnormal components. The feature judgment includes: low-frequency heavy-load wear characteristics of the reducer or high-frequency electromagnetic abnormal characteristics of the motor. The second stage of training involves logically aligning the candidate abnormal component set with the fault mapping relationship set of the constraint operators in the complete structured prompt text based on the large language model. If the candidate abnormal component set matches the fault type in the fault mapping relationship set, a system-level fault category label or a normal label is output. If the candidate abnormal component set does not match the fault type in the fault mapping relationship set, a normal label is output.

6. The industrial robot fault diagnosis method based on segmented context enhancement and logic decoupling according to claim 5, characterized in that, Obtaining the candidate anomaly component set includes: ； in, For the set of candidate exception components, This indicates a decelerator malfunction in joints 1-6. This indicates a motor malfunction in joints 1-6.

7. The industrial robot fault diagnosis method based on segmented context enhancement and logic decoupling according to claim 6, characterized in that, The output of the system-level fault category labels includes: ； in, This is a system-level fault category label or a normal label.

Images